Text Generation
Transformers
Safetensors
PEFT
English
Chinese
qwen3_5
image-text-to-text
veriloop
veriloop-coder
code
coding-agent
software-engineering
repository-understanding
tool-use
lora
harness-engineering
evidence-binding
rollback
uncertainty-calibration
long-context
open-weights
conversational
Upload uncertainty_qwen36_rootfix_run1 contents into uncertainty_adapter
Browse files- .gitattributes +1 -0
- uncertainty_adapter/adapter/README.md +203 -0
- uncertainty_adapter/adapter/adapter_config.json +43 -0
- uncertainty_adapter/adapter/adapter_model.safetensors +3 -0
- uncertainty_adapter/best_checkpoint_manifest.json +33 -0
- uncertainty_adapter/epoch_history.json +140 -0
- uncertainty_adapter/host_manifest.json +148 -0
- uncertainty_adapter/tokenizer/chat_template.jinja +154 -0
- uncertainty_adapter/tokenizer/tokenizer.json +3 -0
- uncertainty_adapter/tokenizer/tokenizer_config.json +31 -0
- uncertainty_adapter/uncertainty_adapter_plan.json +469 -0
- uncertainty_adapter/uncertainty_adapter_train_result.json +564 -0
- uncertainty_adapter/uncertainty_eval.jsonl +55 -0
- uncertainty_adapter/uncertainty_head.pt +3 -0
- uncertainty_adapter/uncertainty_train.jsonl +0 -0
- uncertainty_adapter/uncertainty_training_manifest.json +105 -0
.gitattributes
CHANGED
|
@@ -35,3 +35,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
| 37 |
toolspec_adapter/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
| 37 |
toolspec_adapter/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
uncertainty_adapter/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
uncertainty_adapter/adapter/README.md
ADDED
|
@@ -0,0 +1,203 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: peft
|
| 3 |
+
tags:
|
| 4 |
+
- lora
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
# Model Card for Model ID
|
| 8 |
+
|
| 9 |
+
<!-- Provide a quick summary of what the model is/does. -->
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
## Model Details
|
| 14 |
+
|
| 15 |
+
### Model Description
|
| 16 |
+
|
| 17 |
+
<!-- Provide a longer summary of what this model is. -->
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
- **Developed by:** [More Information Needed]
|
| 22 |
+
- **Funded by [optional]:** [More Information Needed]
|
| 23 |
+
- **Shared by [optional]:** [More Information Needed]
|
| 24 |
+
- **Model type:** [More Information Needed]
|
| 25 |
+
- **Language(s) (NLP):** [More Information Needed]
|
| 26 |
+
- **License:** [More Information Needed]
|
| 27 |
+
- **Finetuned from model [optional]:** [More Information Needed]
|
| 28 |
+
|
| 29 |
+
### Model Sources [optional]
|
| 30 |
+
|
| 31 |
+
<!-- Provide the basic links for the model. -->
|
| 32 |
+
|
| 33 |
+
- **Repository:** [More Information Needed]
|
| 34 |
+
- **Paper [optional]:** [More Information Needed]
|
| 35 |
+
- **Demo [optional]:** [More Information Needed]
|
| 36 |
+
|
| 37 |
+
## Uses
|
| 38 |
+
|
| 39 |
+
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
|
| 40 |
+
|
| 41 |
+
### Direct Use
|
| 42 |
+
|
| 43 |
+
<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
|
| 44 |
+
|
| 45 |
+
[More Information Needed]
|
| 46 |
+
|
| 47 |
+
### Downstream Use [optional]
|
| 48 |
+
|
| 49 |
+
<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
|
| 50 |
+
|
| 51 |
+
[More Information Needed]
|
| 52 |
+
|
| 53 |
+
### Out-of-Scope Use
|
| 54 |
+
|
| 55 |
+
<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
|
| 56 |
+
|
| 57 |
+
[More Information Needed]
|
| 58 |
+
|
| 59 |
+
## Bias, Risks, and Limitations
|
| 60 |
+
|
| 61 |
+
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
|
| 62 |
+
|
| 63 |
+
[More Information Needed]
|
| 64 |
+
|
| 65 |
+
### Recommendations
|
| 66 |
+
|
| 67 |
+
<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
|
| 68 |
+
|
| 69 |
+
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
|
| 70 |
+
|
| 71 |
+
## How to Get Started with the Model
|
| 72 |
+
|
| 73 |
+
Use the code below to get started with the model.
|
| 74 |
+
|
| 75 |
+
[More Information Needed]
|
| 76 |
+
|
| 77 |
+
## Training Details
|
| 78 |
+
|
| 79 |
+
### Training Data
|
| 80 |
+
|
| 81 |
+
<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
|
| 82 |
+
|
| 83 |
+
[More Information Needed]
|
| 84 |
+
|
| 85 |
+
### Training Procedure
|
| 86 |
+
|
| 87 |
+
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
|
| 88 |
+
|
| 89 |
+
#### Preprocessing [optional]
|
| 90 |
+
|
| 91 |
+
[More Information Needed]
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
#### Training Hyperparameters
|
| 95 |
+
|
| 96 |
+
- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
|
| 97 |
+
|
| 98 |
+
#### Speeds, Sizes, Times [optional]
|
| 99 |
+
|
| 100 |
+
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
|
| 101 |
+
|
| 102 |
+
[More Information Needed]
|
| 103 |
+
|
| 104 |
+
## Evaluation
|
| 105 |
+
|
| 106 |
+
<!-- This section describes the evaluation protocols and provides the results. -->
|
| 107 |
+
|
| 108 |
+
### Testing Data, Factors & Metrics
|
| 109 |
+
|
| 110 |
+
#### Testing Data
|
| 111 |
+
|
| 112 |
+
<!-- This should link to a Dataset Card if possible. -->
|
| 113 |
+
|
| 114 |
+
[More Information Needed]
|
| 115 |
+
|
| 116 |
+
#### Factors
|
| 117 |
+
|
| 118 |
+
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
|
| 119 |
+
|
| 120 |
+
[More Information Needed]
|
| 121 |
+
|
| 122 |
+
#### Metrics
|
| 123 |
+
|
| 124 |
+
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
|
| 125 |
+
|
| 126 |
+
[More Information Needed]
|
| 127 |
+
|
| 128 |
+
### Results
|
| 129 |
+
|
| 130 |
+
[More Information Needed]
|
| 131 |
+
|
| 132 |
+
#### Summary
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
|
| 136 |
+
## Model Examination [optional]
|
| 137 |
+
|
| 138 |
+
<!-- Relevant interpretability work for the model goes here -->
|
| 139 |
+
|
| 140 |
+
[More Information Needed]
|
| 141 |
+
|
| 142 |
+
## Environmental Impact
|
| 143 |
+
|
| 144 |
+
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
|
| 145 |
+
|
| 146 |
+
Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
|
| 147 |
+
|
| 148 |
+
- **Hardware Type:** [More Information Needed]
|
| 149 |
+
- **Hours used:** [More Information Needed]
|
| 150 |
+
- **Cloud Provider:** [More Information Needed]
|
| 151 |
+
- **Compute Region:** [More Information Needed]
|
| 152 |
+
- **Carbon Emitted:** [More Information Needed]
|
| 153 |
+
|
| 154 |
+
## Technical Specifications [optional]
|
| 155 |
+
|
| 156 |
+
### Model Architecture and Objective
|
| 157 |
+
|
| 158 |
+
[More Information Needed]
|
| 159 |
+
|
| 160 |
+
### Compute Infrastructure
|
| 161 |
+
|
| 162 |
+
[More Information Needed]
|
| 163 |
+
|
| 164 |
+
#### Hardware
|
| 165 |
+
|
| 166 |
+
[More Information Needed]
|
| 167 |
+
|
| 168 |
+
#### Software
|
| 169 |
+
|
| 170 |
+
[More Information Needed]
|
| 171 |
+
|
| 172 |
+
## Citation [optional]
|
| 173 |
+
|
| 174 |
+
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
|
| 175 |
+
|
| 176 |
+
**BibTeX:**
|
| 177 |
+
|
| 178 |
+
[More Information Needed]
|
| 179 |
+
|
| 180 |
+
**APA:**
|
| 181 |
+
|
| 182 |
+
[More Information Needed]
|
| 183 |
+
|
| 184 |
+
## Glossary [optional]
|
| 185 |
+
|
| 186 |
+
<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
|
| 187 |
+
|
| 188 |
+
[More Information Needed]
|
| 189 |
+
|
| 190 |
+
## More Information [optional]
|
| 191 |
+
|
| 192 |
+
[More Information Needed]
|
| 193 |
+
|
| 194 |
+
## Model Card Authors [optional]
|
| 195 |
+
|
| 196 |
+
[More Information Needed]
|
| 197 |
+
|
| 198 |
+
## Model Card Contact
|
| 199 |
+
|
| 200 |
+
[More Information Needed]
|
| 201 |
+
### Framework versions
|
| 202 |
+
|
| 203 |
+
- PEFT 0.19.0
|
uncertainty_adapter/adapter/adapter_config.json
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alora_invocation_tokens": null,
|
| 3 |
+
"alpha_pattern": {},
|
| 4 |
+
"arrow_config": null,
|
| 5 |
+
"auto_mapping": null,
|
| 6 |
+
"base_model_name_or_path": null,
|
| 7 |
+
"bias": "none",
|
| 8 |
+
"corda_config": null,
|
| 9 |
+
"ensure_weight_tying": false,
|
| 10 |
+
"eva_config": null,
|
| 11 |
+
"exclude_modules": null,
|
| 12 |
+
"fan_in_fan_out": false,
|
| 13 |
+
"inference_mode": true,
|
| 14 |
+
"init_lora_weights": true,
|
| 15 |
+
"layer_replication": null,
|
| 16 |
+
"layers_pattern": null,
|
| 17 |
+
"layers_to_transform": null,
|
| 18 |
+
"loftq_config": {},
|
| 19 |
+
"lora_alpha": 16,
|
| 20 |
+
"lora_bias": false,
|
| 21 |
+
"lora_dropout": 0.05,
|
| 22 |
+
"lora_ga_config": null,
|
| 23 |
+
"megatron_config": null,
|
| 24 |
+
"megatron_core": "megatron.core",
|
| 25 |
+
"modules_to_save": null,
|
| 26 |
+
"peft_type": "LORA",
|
| 27 |
+
"peft_version": "0.19.0",
|
| 28 |
+
"qalora_group_size": 16,
|
| 29 |
+
"r": 8,
|
| 30 |
+
"rank_pattern": {},
|
| 31 |
+
"revision": null,
|
| 32 |
+
"target_modules": [
|
| 33 |
+
"surface_host.uncertainty_head.proj",
|
| 34 |
+
"surface_host.uncertainty_head.calibration_mlp"
|
| 35 |
+
],
|
| 36 |
+
"target_parameters": null,
|
| 37 |
+
"task_type": "FEATURE_EXTRACTION",
|
| 38 |
+
"trainable_token_indices": null,
|
| 39 |
+
"use_bdlora": null,
|
| 40 |
+
"use_dora": false,
|
| 41 |
+
"use_qalora": false,
|
| 42 |
+
"use_rslora": false
|
| 43 |
+
}
|
uncertainty_adapter/adapter/adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fcca67011c0b0b0eaa44fc4826f4908bdd6ca8d82b192916a6e35556e292fa7f
|
| 3 |
+
size 98968
|
uncertainty_adapter/best_checkpoint_manifest.json
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"best_epoch": 2,
|
| 3 |
+
"best_quality_score": 0.8625935807221907,
|
| 4 |
+
"eval_metrics": {
|
| 5 |
+
"count": 55,
|
| 6 |
+
"mae": {
|
| 7 |
+
"u_answer": 0.15174226462841034,
|
| 8 |
+
"u_evidence": 0.19610758125782013,
|
| 9 |
+
"u_exec": 0.18561214208602905,
|
| 10 |
+
"u_risk": 0.1553734689950943,
|
| 11 |
+
"u_spec": 0.21633382141590118
|
| 12 |
+
},
|
| 13 |
+
"mean_mae": 0.18103384971618652,
|
| 14 |
+
"mean_rmse": 0.24169571697711945,
|
| 15 |
+
"moderate_accuracy": 0.6727272727272727,
|
| 16 |
+
"quality_score": 0.8625935807221907,
|
| 17 |
+
"rmse": {
|
| 18 |
+
"u_answer": 0.18724055588245392,
|
| 19 |
+
"u_evidence": 0.22527915239334106,
|
| 20 |
+
"u_exec": 0.25238174200057983,
|
| 21 |
+
"u_risk": 0.20667441189289093,
|
| 22 |
+
"u_spec": 0.3369026482105255
|
| 23 |
+
},
|
| 24 |
+
"tight_accuracy": 0.4,
|
| 25 |
+
"weighted_mae": 0.18083095811830807,
|
| 26 |
+
"weighted_rmse": 0.24125460771003793
|
| 27 |
+
},
|
| 28 |
+
"train_metrics": {
|
| 29 |
+
"loss": 0.009006613283418119,
|
| 30 |
+
"micro_batches": 154,
|
| 31 |
+
"optimizer_steps": 10
|
| 32 |
+
}
|
| 33 |
+
}
|
uncertainty_adapter/epoch_history.json
ADDED
|
@@ -0,0 +1,140 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"epochs": [
|
| 3 |
+
{
|
| 4 |
+
"epoch": 1,
|
| 5 |
+
"eval_metrics": {
|
| 6 |
+
"count": 55,
|
| 7 |
+
"mae": {
|
| 8 |
+
"u_answer": 0.1848176270723343,
|
| 9 |
+
"u_evidence": 0.20735961198806763,
|
| 10 |
+
"u_exec": 0.21166233718395233,
|
| 11 |
+
"u_risk": 0.1652287244796753,
|
| 12 |
+
"u_spec": 0.24522197246551514
|
| 13 |
+
},
|
| 14 |
+
"mean_mae": 0.2028580605983734,
|
| 15 |
+
"mean_rmse": 0.2640994191169739,
|
| 16 |
+
"moderate_accuracy": 0.6036363636363636,
|
| 17 |
+
"quality_score": 0.8344919736975306,
|
| 18 |
+
"rmse": {
|
| 19 |
+
"u_answer": 0.2176007330417633,
|
| 20 |
+
"u_evidence": 0.23360508680343628,
|
| 21 |
+
"u_exec": 0.2795017957687378,
|
| 22 |
+
"u_risk": 0.22334341704845428,
|
| 23 |
+
"u_spec": 0.36644604802131653
|
| 24 |
+
},
|
| 25 |
+
"tight_accuracy": 0.36727272727272725,
|
| 26 |
+
"weighted_mae": 0.20175925098418107,
|
| 27 |
+
"weighted_rmse": 0.26307498343893954
|
| 28 |
+
},
|
| 29 |
+
"improved": true,
|
| 30 |
+
"quality_score": 0.8344919736975306,
|
| 31 |
+
"train_metrics": {
|
| 32 |
+
"loss": 0.011121059906239584,
|
| 33 |
+
"micro_batches": 154,
|
| 34 |
+
"optimizer_steps": 10
|
| 35 |
+
}
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"epoch": 2,
|
| 39 |
+
"eval_metrics": {
|
| 40 |
+
"count": 55,
|
| 41 |
+
"mae": {
|
| 42 |
+
"u_answer": 0.15174226462841034,
|
| 43 |
+
"u_evidence": 0.19610758125782013,
|
| 44 |
+
"u_exec": 0.18561214208602905,
|
| 45 |
+
"u_risk": 0.1553734689950943,
|
| 46 |
+
"u_spec": 0.21633382141590118
|
| 47 |
+
},
|
| 48 |
+
"mean_mae": 0.18103384971618652,
|
| 49 |
+
"mean_rmse": 0.24169571697711945,
|
| 50 |
+
"moderate_accuracy": 0.6727272727272727,
|
| 51 |
+
"quality_score": 0.8625935807221907,
|
| 52 |
+
"rmse": {
|
| 53 |
+
"u_answer": 0.18724055588245392,
|
| 54 |
+
"u_evidence": 0.22527915239334106,
|
| 55 |
+
"u_exec": 0.25238174200057983,
|
| 56 |
+
"u_risk": 0.20667441189289093,
|
| 57 |
+
"u_spec": 0.3369026482105255
|
| 58 |
+
},
|
| 59 |
+
"tight_accuracy": 0.4,
|
| 60 |
+
"weighted_mae": 0.18083095811830807,
|
| 61 |
+
"weighted_rmse": 0.24125460771003793
|
| 62 |
+
},
|
| 63 |
+
"improved": true,
|
| 64 |
+
"quality_score": 0.8625935807221907,
|
| 65 |
+
"train_metrics": {
|
| 66 |
+
"loss": 0.009006613283418119,
|
| 67 |
+
"micro_batches": 154,
|
| 68 |
+
"optimizer_steps": 10
|
| 69 |
+
}
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"epoch": 3,
|
| 73 |
+
"eval_metrics": {
|
| 74 |
+
"count": 55,
|
| 75 |
+
"mae": {
|
| 76 |
+
"u_answer": 0.14970968663692474,
|
| 77 |
+
"u_evidence": 0.19498984515666962,
|
| 78 |
+
"u_exec": 0.18136857450008392,
|
| 79 |
+
"u_risk": 0.1565144807100296,
|
| 80 |
+
"u_spec": 0.21266111731529236
|
| 81 |
+
},
|
| 82 |
+
"mean_mae": 0.17904874682426453,
|
| 83 |
+
"mean_rmse": 0.23914317786693573,
|
| 84 |
+
"moderate_accuracy": 0.6545454545454545,
|
| 85 |
+
"quality_score": 0.8630735522772117,
|
| 86 |
+
"rmse": {
|
| 87 |
+
"u_answer": 0.1858142465353012,
|
| 88 |
+
"u_evidence": 0.22504529356956482,
|
| 89 |
+
"u_exec": 0.24751117825508118,
|
| 90 |
+
"u_risk": 0.20870783925056458,
|
| 91 |
+
"u_spec": 0.3286373019218445
|
| 92 |
+
},
|
| 93 |
+
"tight_accuracy": 0.3927272727272727,
|
| 94 |
+
"weighted_mae": 0.1789679336398722,
|
| 95 |
+
"weighted_rmse": 0.23888496180570432
|
| 96 |
+
},
|
| 97 |
+
"improved": false,
|
| 98 |
+
"quality_score": 0.8630735522772117,
|
| 99 |
+
"train_metrics": {
|
| 100 |
+
"loss": 0.008164266918108544,
|
| 101 |
+
"micro_batches": 154,
|
| 102 |
+
"optimizer_steps": 10
|
| 103 |
+
}
|
| 104 |
+
},
|
| 105 |
+
{
|
| 106 |
+
"epoch": 4,
|
| 107 |
+
"eval_metrics": {
|
| 108 |
+
"count": 55,
|
| 109 |
+
"mae": {
|
| 110 |
+
"u_answer": 0.1496635228395462,
|
| 111 |
+
"u_evidence": 0.19500812888145447,
|
| 112 |
+
"u_exec": 0.18158338963985443,
|
| 113 |
+
"u_risk": 0.1564789116382599,
|
| 114 |
+
"u_spec": 0.21253055334091187
|
| 115 |
+
},
|
| 116 |
+
"mean_mae": 0.1790529191493988,
|
| 117 |
+
"mean_rmse": 0.23915143311023712,
|
| 118 |
+
"moderate_accuracy": 0.6545454545454545,
|
| 119 |
+
"quality_score": 0.8630639342548476,
|
| 120 |
+
"rmse": {
|
| 121 |
+
"u_answer": 0.1857716143131256,
|
| 122 |
+
"u_evidence": 0.22511473298072815,
|
| 123 |
+
"u_exec": 0.24768365919589996,
|
| 124 |
+
"u_risk": 0.20870527625083923,
|
| 125 |
+
"u_spec": 0.32848188281059265
|
| 126 |
+
},
|
| 127 |
+
"tight_accuracy": 0.3927272727272727,
|
| 128 |
+
"weighted_mae": 0.17897483332915498,
|
| 129 |
+
"weighted_rmse": 0.2388972981611497
|
| 130 |
+
},
|
| 131 |
+
"improved": false,
|
| 132 |
+
"quality_score": 0.8630639342548476,
|
| 133 |
+
"train_metrics": {
|
| 134 |
+
"loss": 0.008081083455086723,
|
| 135 |
+
"micro_batches": 154,
|
| 136 |
+
"optimizer_steps": 10
|
| 137 |
+
}
|
| 138 |
+
}
|
| 139 |
+
]
|
| 140 |
+
}
|
uncertainty_adapter/host_manifest.json
ADDED
|
@@ -0,0 +1,148 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hidden_size": 2048,
|
| 3 |
+
"host_config": {
|
| 4 |
+
"attn_implementation": null,
|
| 5 |
+
"backbone_name_or_path": "/public/wang_libo/veriloop_coder_e1/model",
|
| 6 |
+
"device_map": null,
|
| 7 |
+
"dtype": null,
|
| 8 |
+
"evidence_rank_hint": 8,
|
| 9 |
+
"expose_backbone_inventory": false,
|
| 10 |
+
"freeze_backbone": true,
|
| 11 |
+
"hidden_size_override": 2048,
|
| 12 |
+
"host_dropout": 0.0,
|
| 13 |
+
"identity_rank_hint": 8,
|
| 14 |
+
"load_backbone_weights": false,
|
| 15 |
+
"local_files_only": true,
|
| 16 |
+
"low_cpu_mem_usage": true,
|
| 17 |
+
"memory_rank_hint": 4,
|
| 18 |
+
"rollback_rank_hint": 8,
|
| 19 |
+
"runtime_rank_hint": 8,
|
| 20 |
+
"toolspec_rank_hint": 8,
|
| 21 |
+
"trust_remote_code": true,
|
| 22 |
+
"uncertainty_rank_hint": 8,
|
| 23 |
+
"use_safetensors": null,
|
| 24 |
+
"validator_rank_hint": 8
|
| 25 |
+
},
|
| 26 |
+
"load_result": {
|
| 27 |
+
"has_base_config": true,
|
| 28 |
+
"has_base_model": true,
|
| 29 |
+
"hidden_size": 2048,
|
| 30 |
+
"notes": [
|
| 31 |
+
"class=AutoModelForCausalLM",
|
| 32 |
+
"quant=4bit"
|
| 33 |
+
],
|
| 34 |
+
"source": "trainer_qwen36_loader"
|
| 35 |
+
},
|
| 36 |
+
"peft_named_modules": [
|
| 37 |
+
"claim_extractor",
|
| 38 |
+
"claim_extractor.adapter",
|
| 39 |
+
"dropout",
|
| 40 |
+
"episodic_memory",
|
| 41 |
+
"episodic_memory.adapter",
|
| 42 |
+
"evidence_binding",
|
| 43 |
+
"evidence_binding.adapter",
|
| 44 |
+
"failure_signal_bridge",
|
| 45 |
+
"failure_signal_bridge.rollback_bridge",
|
| 46 |
+
"identity_adapter",
|
| 47 |
+
"identity_adapter.bridge",
|
| 48 |
+
"identity_guard",
|
| 49 |
+
"identity_guard.adapter",
|
| 50 |
+
"input_norm",
|
| 51 |
+
"memory_boundary_guard",
|
| 52 |
+
"memory_boundary_guard.adapter",
|
| 53 |
+
"memory_boundary_guard.rollback_filter",
|
| 54 |
+
"patch_binding",
|
| 55 |
+
"patch_binding.adapter",
|
| 56 |
+
"permission_context_manager",
|
| 57 |
+
"permission_context_manager.adapter",
|
| 58 |
+
"progress_state_tracker",
|
| 59 |
+
"progress_state_tracker.adapter",
|
| 60 |
+
"progress_state_tracker.rollback_memory",
|
| 61 |
+
"proof_carrying_hints",
|
| 62 |
+
"proof_carrying_hints.bridge",
|
| 63 |
+
"provenance_binding",
|
| 64 |
+
"provenance_binding.adapter",
|
| 65 |
+
"public_identity_head",
|
| 66 |
+
"public_identity_head.proj",
|
| 67 |
+
"query_runtime_engine",
|
| 68 |
+
"query_runtime_engine.adapter",
|
| 69 |
+
"request_normalizer",
|
| 70 |
+
"request_normalizer.adapter",
|
| 71 |
+
"rollback_adapter",
|
| 72 |
+
"rollback_adapter.head",
|
| 73 |
+
"rollback_engine",
|
| 74 |
+
"rollback_engine.adapter",
|
| 75 |
+
"runtime_binding",
|
| 76 |
+
"runtime_binding.adapter",
|
| 77 |
+
"runtime_harness_adapter",
|
| 78 |
+
"runtime_harness_adapter.bridge",
|
| 79 |
+
"runtime_harness_uncertainty_bridge",
|
| 80 |
+
"runtime_harness_uncertainty_bridge.adapter",
|
| 81 |
+
"sandbox_rollback_bridge",
|
| 82 |
+
"sandbox_rollback_bridge.adapter",
|
| 83 |
+
"session_compactor",
|
| 84 |
+
"session_compactor.adapter",
|
| 85 |
+
"session_state_manager",
|
| 86 |
+
"session_state_manager.adapter",
|
| 87 |
+
"session_state_manager.rollback_state",
|
| 88 |
+
"tool_protocol_adapter",
|
| 89 |
+
"tool_protocol_adapter.bridge",
|
| 90 |
+
"toolspec_bridge",
|
| 91 |
+
"toolspec_bridge.adapter",
|
| 92 |
+
"toolspec_head",
|
| 93 |
+
"toolspec_head.param_schema_adapter",
|
| 94 |
+
"toolspec_head.postcondition_adapter",
|
| 95 |
+
"toolspec_head.precondition_adapter",
|
| 96 |
+
"toolspec_head.receipt_formatter",
|
| 97 |
+
"toolspec_head.trigger_gate",
|
| 98 |
+
"toolspec_head.validator_gate",
|
| 99 |
+
"uncertainty_head",
|
| 100 |
+
"uncertainty_head.calibration_mlp",
|
| 101 |
+
"uncertainty_head.calibration_mlp.base_layer",
|
| 102 |
+
"uncertainty_head.calibration_mlp.lora_A",
|
| 103 |
+
"uncertainty_head.calibration_mlp.lora_A.default",
|
| 104 |
+
"uncertainty_head.calibration_mlp.lora_B",
|
| 105 |
+
"uncertainty_head.calibration_mlp.lora_B.default",
|
| 106 |
+
"uncertainty_head.calibration_mlp.lora_dropout",
|
| 107 |
+
"uncertainty_head.calibration_mlp.lora_dropout.default",
|
| 108 |
+
"uncertainty_head.calibration_mlp.lora_embedding_A",
|
| 109 |
+
"uncertainty_head.calibration_mlp.lora_embedding_B",
|
| 110 |
+
"uncertainty_head.calibration_mlp.lora_magnitude_vector",
|
| 111 |
+
"uncertainty_head.proj",
|
| 112 |
+
"uncertainty_head.proj.base_layer",
|
| 113 |
+
"uncertainty_head.proj.lora_A",
|
| 114 |
+
"uncertainty_head.proj.lora_A.default",
|
| 115 |
+
"uncertainty_head.proj.lora_B",
|
| 116 |
+
"uncertainty_head.proj.lora_B.default",
|
| 117 |
+
"uncertainty_head.proj.lora_dropout",
|
| 118 |
+
"uncertainty_head.proj.lora_dropout.default",
|
| 119 |
+
"uncertainty_head.proj.lora_embedding_A",
|
| 120 |
+
"uncertainty_head.proj.lora_embedding_B",
|
| 121 |
+
"uncertainty_head.proj.lora_magnitude_vector",
|
| 122 |
+
"validator_feedback_bridge",
|
| 123 |
+
"validator_feedback_bridge.adapter",
|
| 124 |
+
"validator_feedback_loop",
|
| 125 |
+
"validator_feedback_loop.rollback_adapter",
|
| 126 |
+
"validator_receipt_bridge",
|
| 127 |
+
"validator_receipt_bridge.adapter",
|
| 128 |
+
"validator_uncertainty_bridge",
|
| 129 |
+
"validator_uncertainty_bridge.adapter",
|
| 130 |
+
"workspace_snapshot_manager",
|
| 131 |
+
"workspace_snapshot_manager.rollback_context",
|
| 132 |
+
"worktree_binding",
|
| 133 |
+
"worktree_binding.adapter",
|
| 134 |
+
"worktree_manager",
|
| 135 |
+
"worktree_manager.adapter"
|
| 136 |
+
],
|
| 137 |
+
"trainable_parameter_report": {
|
| 138 |
+
"backbone_frozen": true,
|
| 139 |
+
"backbone_present": true,
|
| 140 |
+
"hidden_size": 2048,
|
| 141 |
+
"host_parameters": 176310317,
|
| 142 |
+
"host_trainable_parameters": 49192,
|
| 143 |
+
"total_parameters": 34132368045,
|
| 144 |
+
"trainable_parameters": 49192,
|
| 145 |
+
"version": "veriloop.coder_peft_host.v1"
|
| 146 |
+
},
|
| 147 |
+
"version": "veriloop.coder_peft_host.v1"
|
| 148 |
+
}
|
uncertainty_adapter/tokenizer/chat_template.jinja
ADDED
|
@@ -0,0 +1,154 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{%- set image_count = namespace(value=0) %}
|
| 2 |
+
{%- set video_count = namespace(value=0) %}
|
| 3 |
+
{%- macro render_content(content, do_vision_count, is_system_content=false) %}
|
| 4 |
+
{%- if content is string %}
|
| 5 |
+
{{- content }}
|
| 6 |
+
{%- elif content is iterable and content is not mapping %}
|
| 7 |
+
{%- for item in content %}
|
| 8 |
+
{%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
|
| 9 |
+
{%- if is_system_content %}
|
| 10 |
+
{{- raise_exception('System message cannot contain images.') }}
|
| 11 |
+
{%- endif %}
|
| 12 |
+
{%- if do_vision_count %}
|
| 13 |
+
{%- set image_count.value = image_count.value + 1 %}
|
| 14 |
+
{%- endif %}
|
| 15 |
+
{%- if add_vision_id %}
|
| 16 |
+
{{- 'Picture ' ~ image_count.value ~ ': ' }}
|
| 17 |
+
{%- endif %}
|
| 18 |
+
{{- '<|vision_start|><|image_pad|><|vision_end|>' }}
|
| 19 |
+
{%- elif 'video' in item or item.type == 'video' %}
|
| 20 |
+
{%- if is_system_content %}
|
| 21 |
+
{{- raise_exception('System message cannot contain videos.') }}
|
| 22 |
+
{%- endif %}
|
| 23 |
+
{%- if do_vision_count %}
|
| 24 |
+
{%- set video_count.value = video_count.value + 1 %}
|
| 25 |
+
{%- endif %}
|
| 26 |
+
{%- if add_vision_id %}
|
| 27 |
+
{{- 'Video ' ~ video_count.value ~ ': ' }}
|
| 28 |
+
{%- endif %}
|
| 29 |
+
{{- '<|vision_start|><|video_pad|><|vision_end|>' }}
|
| 30 |
+
{%- elif 'text' in item %}
|
| 31 |
+
{{- item.text }}
|
| 32 |
+
{%- else %}
|
| 33 |
+
{{- raise_exception('Unexpected item type in content.') }}
|
| 34 |
+
{%- endif %}
|
| 35 |
+
{%- endfor %}
|
| 36 |
+
{%- elif content is none or content is undefined %}
|
| 37 |
+
{{- '' }}
|
| 38 |
+
{%- else %}
|
| 39 |
+
{{- raise_exception('Unexpected content type.') }}
|
| 40 |
+
{%- endif %}
|
| 41 |
+
{%- endmacro %}
|
| 42 |
+
{%- if not messages %}
|
| 43 |
+
{{- raise_exception('No messages provided.') }}
|
| 44 |
+
{%- endif %}
|
| 45 |
+
{%- if tools and tools is iterable and tools is not mapping %}
|
| 46 |
+
{{- '<|im_start|>system\n' }}
|
| 47 |
+
{{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
|
| 48 |
+
{%- for tool in tools %}
|
| 49 |
+
{{- "\n" }}
|
| 50 |
+
{{- tool | tojson }}
|
| 51 |
+
{%- endfor %}
|
| 52 |
+
{{- "\n</tools>" }}
|
| 53 |
+
{{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
|
| 54 |
+
{%- if messages[0].role == 'system' %}
|
| 55 |
+
{%- set content = render_content(messages[0].content, false, true)|trim %}
|
| 56 |
+
{%- if content %}
|
| 57 |
+
{{- '\n\n' + content }}
|
| 58 |
+
{%- endif %}
|
| 59 |
+
{%- endif %}
|
| 60 |
+
{{- '<|im_end|>\n' }}
|
| 61 |
+
{%- else %}
|
| 62 |
+
{%- if messages[0].role == 'system' %}
|
| 63 |
+
{%- set content = render_content(messages[0].content, false, true)|trim %}
|
| 64 |
+
{{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
|
| 65 |
+
{%- endif %}
|
| 66 |
+
{%- endif %}
|
| 67 |
+
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
|
| 68 |
+
{%- for message in messages[::-1] %}
|
| 69 |
+
{%- set index = (messages|length - 1) - loop.index0 %}
|
| 70 |
+
{%- if ns.multi_step_tool and message.role == "user" %}
|
| 71 |
+
{%- set content = render_content(message.content, false)|trim %}
|
| 72 |
+
{%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
|
| 73 |
+
{%- set ns.multi_step_tool = false %}
|
| 74 |
+
{%- set ns.last_query_index = index %}
|
| 75 |
+
{%- endif %}
|
| 76 |
+
{%- endif %}
|
| 77 |
+
{%- endfor %}
|
| 78 |
+
{%- if ns.multi_step_tool %}
|
| 79 |
+
{{- raise_exception('No user query found in messages.') }}
|
| 80 |
+
{%- endif %}
|
| 81 |
+
{%- for message in messages %}
|
| 82 |
+
{%- set content = render_content(message.content, true)|trim %}
|
| 83 |
+
{%- if message.role == "system" %}
|
| 84 |
+
{%- if not loop.first %}
|
| 85 |
+
{{- raise_exception('System message must be at the beginning.') }}
|
| 86 |
+
{%- endif %}
|
| 87 |
+
{%- elif message.role == "user" %}
|
| 88 |
+
{{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
|
| 89 |
+
{%- elif message.role == "assistant" %}
|
| 90 |
+
{%- set reasoning_content = '' %}
|
| 91 |
+
{%- if message.reasoning_content is string %}
|
| 92 |
+
{%- set reasoning_content = message.reasoning_content %}
|
| 93 |
+
{%- else %}
|
| 94 |
+
{%- if '</think>' in content %}
|
| 95 |
+
{%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
|
| 96 |
+
{%- set content = content.split('</think>')[-1].lstrip('\n') %}
|
| 97 |
+
{%- endif %}
|
| 98 |
+
{%- endif %}
|
| 99 |
+
{%- set reasoning_content = reasoning_content|trim %}
|
| 100 |
+
{%- if (preserve_thinking is defined and preserve_thinking is true) or (loop.index0 > ns.last_query_index) %}
|
| 101 |
+
{{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
|
| 102 |
+
{%- else %}
|
| 103 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 104 |
+
{%- endif %}
|
| 105 |
+
{%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
|
| 106 |
+
{%- for tool_call in message.tool_calls %}
|
| 107 |
+
{%- if tool_call.function is defined %}
|
| 108 |
+
{%- set tool_call = tool_call.function %}
|
| 109 |
+
{%- endif %}
|
| 110 |
+
{%- if loop.first %}
|
| 111 |
+
{%- if content|trim %}
|
| 112 |
+
{{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
|
| 113 |
+
{%- else %}
|
| 114 |
+
{{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
|
| 115 |
+
{%- endif %}
|
| 116 |
+
{%- else %}
|
| 117 |
+
{{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
|
| 118 |
+
{%- endif %}
|
| 119 |
+
{%- if tool_call.arguments is defined %}
|
| 120 |
+
{%- for args_name, args_value in tool_call.arguments|items %}
|
| 121 |
+
{{- '<parameter=' + args_name + '>\n' }}
|
| 122 |
+
{%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}
|
| 123 |
+
{{- args_value }}
|
| 124 |
+
{{- '\n</parameter>\n' }}
|
| 125 |
+
{%- endfor %}
|
| 126 |
+
{%- endif %}
|
| 127 |
+
{{- '</function>\n</tool_call>' }}
|
| 128 |
+
{%- endfor %}
|
| 129 |
+
{%- endif %}
|
| 130 |
+
{{- '<|im_end|>\n' }}
|
| 131 |
+
{%- elif message.role == "tool" %}
|
| 132 |
+
{%- if loop.previtem and loop.previtem.role != "tool" %}
|
| 133 |
+
{{- '<|im_start|>user' }}
|
| 134 |
+
{%- endif %}
|
| 135 |
+
{{- '\n<tool_response>\n' }}
|
| 136 |
+
{{- content }}
|
| 137 |
+
{{- '\n</tool_response>' }}
|
| 138 |
+
{%- if not loop.last and loop.nextitem.role != "tool" %}
|
| 139 |
+
{{- '<|im_end|>\n' }}
|
| 140 |
+
{%- elif loop.last %}
|
| 141 |
+
{{- '<|im_end|>\n' }}
|
| 142 |
+
{%- endif %}
|
| 143 |
+
{%- else %}
|
| 144 |
+
{{- raise_exception('Unexpected message role.') }}
|
| 145 |
+
{%- endif %}
|
| 146 |
+
{%- endfor %}
|
| 147 |
+
{%- if add_generation_prompt %}
|
| 148 |
+
{{- '<|im_start|>assistant\n' }}
|
| 149 |
+
{%- if enable_thinking is defined and enable_thinking is false %}
|
| 150 |
+
{{- '<think>\n\n</think>\n\n' }}
|
| 151 |
+
{%- else %}
|
| 152 |
+
{{- '<think>\n' }}
|
| 153 |
+
{%- endif %}
|
| 154 |
+
{%- endif %}
|
uncertainty_adapter/tokenizer/tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ea2e66b594a0906e9a547c9e6ff9e5fb8a8198439c8cf7d6dc48f23529161223
|
| 3 |
+
size 19989442
|
uncertainty_adapter/tokenizer/tokenizer_config.json
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_prefix_space": false,
|
| 3 |
+
"audio_bos_token": "<|audio_start|>",
|
| 4 |
+
"audio_eos_token": "<|audio_end|>",
|
| 5 |
+
"audio_token": "<|audio_pad|>",
|
| 6 |
+
"backend": "tokenizers",
|
| 7 |
+
"bos_token": null,
|
| 8 |
+
"clean_up_tokenization_spaces": false,
|
| 9 |
+
"eos_token": "<|im_end|>",
|
| 10 |
+
"errors": "replace",
|
| 11 |
+
"image_token": "<|image_pad|>",
|
| 12 |
+
"is_local": true,
|
| 13 |
+
"model_max_length": 262144,
|
| 14 |
+
"model_specific_special_tokens": {
|
| 15 |
+
"audio_bos_token": "<|audio_start|>",
|
| 16 |
+
"audio_eos_token": "<|audio_end|>",
|
| 17 |
+
"audio_token": "<|audio_pad|>",
|
| 18 |
+
"image_token": "<|image_pad|>",
|
| 19 |
+
"video_token": "<|video_pad|>",
|
| 20 |
+
"vision_bos_token": "<|vision_start|>",
|
| 21 |
+
"vision_eos_token": "<|vision_end|>"
|
| 22 |
+
},
|
| 23 |
+
"pad_token": "<|endoftext|>",
|
| 24 |
+
"pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
|
| 25 |
+
"split_special_tokens": false,
|
| 26 |
+
"tokenizer_class": "TokenizersBackend",
|
| 27 |
+
"unk_token": null,
|
| 28 |
+
"video_token": "<|video_pad|>",
|
| 29 |
+
"vision_bos_token": "<|vision_start|>",
|
| 30 |
+
"vision_eos_token": "<|vision_end|>"
|
| 31 |
+
}
|
uncertainty_adapter/uncertainty_adapter_plan.json
ADDED
|
@@ -0,0 +1,469 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"adapter_alpha": 16,
|
| 3 |
+
"adapter_dropout": 0.05,
|
| 4 |
+
"adapter_rank": 8,
|
| 5 |
+
"backbone_manifest": {
|
| 6 |
+
"host_inventory_size": 79
|
| 7 |
+
},
|
| 8 |
+
"config": {
|
| 9 |
+
"allow_backbone_bridge": false,
|
| 10 |
+
"backbone": "/public/wang_libo/veriloop_coder_e1/model",
|
| 11 |
+
"bf16": true,
|
| 12 |
+
"cache_dir": null,
|
| 13 |
+
"cpu_max_memory_gib": 96,
|
| 14 |
+
"dataset_jsonl": null,
|
| 15 |
+
"dim_weights": {
|
| 16 |
+
"u_answer": 1.1,
|
| 17 |
+
"u_evidence": 1.35,
|
| 18 |
+
"u_exec": 1.35,
|
| 19 |
+
"u_risk": 1.55,
|
| 20 |
+
"u_spec": 1.25
|
| 21 |
+
},
|
| 22 |
+
"early_stopping_min_delta": 0.001,
|
| 23 |
+
"early_stopping_patience": 2,
|
| 24 |
+
"enable_synthetic_dataset": true,
|
| 25 |
+
"eval_jsonl": null,
|
| 26 |
+
"eval_samples_per_mode": 5,
|
| 27 |
+
"fp16": false,
|
| 28 |
+
"gpu_max_memory_gib": 44,
|
| 29 |
+
"gradient_accumulation_steps": 16,
|
| 30 |
+
"host_dropout": 0.0,
|
| 31 |
+
"learning_rate": 2e-05,
|
| 32 |
+
"local_files_only": true,
|
| 33 |
+
"logging_steps": 10,
|
| 34 |
+
"max_grad_norm": 1.0,
|
| 35 |
+
"max_length": 1664,
|
| 36 |
+
"min_epochs_before_early_stop": 2,
|
| 37 |
+
"num_train_epochs": 4.0,
|
| 38 |
+
"output_dir": "./outputs/uncertainty_qwen36_rootfix_run1",
|
| 39 |
+
"per_device_eval_batch_size": 1,
|
| 40 |
+
"per_device_train_batch_size": 1,
|
| 41 |
+
"prefer_best_checkpoint_export": true,
|
| 42 |
+
"probe_dropout": 0.03,
|
| 43 |
+
"probe_hidden_factor": 0.75,
|
| 44 |
+
"product_line": "veriloop_coder",
|
| 45 |
+
"quantization_mode": "4bit",
|
| 46 |
+
"quantization_required": false,
|
| 47 |
+
"require_trainable_targets": true,
|
| 48 |
+
"revision": null,
|
| 49 |
+
"run_post_train_eval": true,
|
| 50 |
+
"save_best_checkpoint": true,
|
| 51 |
+
"seed": 11,
|
| 52 |
+
"selection_mode": "minimal",
|
| 53 |
+
"train_samples_per_mode": 14,
|
| 54 |
+
"training_mode": "mounted_head",
|
| 55 |
+
"trust_remote_code": true,
|
| 56 |
+
"use_double_quant": true,
|
| 57 |
+
"warmup_ratio": 0.05,
|
| 58 |
+
"weight_decay": 0.0,
|
| 59 |
+
"weighted_mae_penalty": 0.5,
|
| 60 |
+
"weighted_rmse_penalty": 0.5
|
| 61 |
+
},
|
| 62 |
+
"dataset_summary": {
|
| 63 |
+
"eval_modes": {
|
| 64 |
+
"conflicting_evidence": 5,
|
| 65 |
+
"evidence_gap": 5,
|
| 66 |
+
"exec_required": 5,
|
| 67 |
+
"high_risk": 5,
|
| 68 |
+
"low_uncertainty": 5,
|
| 69 |
+
"patch_pending": 5,
|
| 70 |
+
"reverse_engineering_ambiguity": 5,
|
| 71 |
+
"self_check_failure": 5,
|
| 72 |
+
"spec_mismatch": 5,
|
| 73 |
+
"validator_negation": 5,
|
| 74 |
+
"worktree_conflict": 5
|
| 75 |
+
},
|
| 76 |
+
"eval_size": 55,
|
| 77 |
+
"train_modes": {
|
| 78 |
+
"conflicting_evidence": 14,
|
| 79 |
+
"evidence_gap": 14,
|
| 80 |
+
"exec_required": 14,
|
| 81 |
+
"high_risk": 14,
|
| 82 |
+
"low_uncertainty": 14,
|
| 83 |
+
"patch_pending": 14,
|
| 84 |
+
"reverse_engineering_ambiguity": 14,
|
| 85 |
+
"self_check_failure": 14,
|
| 86 |
+
"spec_mismatch": 14,
|
| 87 |
+
"validator_negation": 14,
|
| 88 |
+
"worktree_conflict": 14
|
| 89 |
+
},
|
| 90 |
+
"train_size": 154
|
| 91 |
+
},
|
| 92 |
+
"head_strategy": "host_head",
|
| 93 |
+
"notes": [
|
| 94 |
+
"Primary route is host-surface-first uncertainty training.",
|
| 95 |
+
"Validator receipts, self-check failure, reverse-engineering ambiguity, worktree conflicts, and patch continuity are first-class signals.",
|
| 96 |
+
"DualPath, Full AttnRes, mHC, routers, experts, and broad attention-layer PEFT stay excluded.",
|
| 97 |
+
"This adapter should improve runtime uncertainty routing, not general coding free-formity."
|
| 98 |
+
],
|
| 99 |
+
"peft_method": "lora_narrow",
|
| 100 |
+
"recipe": {
|
| 101 |
+
"adapter_family": "uncertainty",
|
| 102 |
+
"backbone": "/public/wang_libo/veriloop_coder_e1/model",
|
| 103 |
+
"backbone_family": "qwen_dense",
|
| 104 |
+
"excluded_patterns": [
|
| 105 |
+
"(?i)\\bdualpath\\b",
|
| 106 |
+
"(?i)\\bmhc\\b",
|
| 107 |
+
"(?i)\\bfull[_\\- ]?attnres\\b",
|
| 108 |
+
"(?i)\\battnres(_full)?\\b",
|
| 109 |
+
"(?i)\\brouter\\b",
|
| 110 |
+
"(?i)\\bexperts?\\b",
|
| 111 |
+
"(?i)\\bmoe\\b.*\\b(gate|router|expert)\\b",
|
| 112 |
+
"(?i)\\brope\\b.*\\b(freq|inv_freq|theta|rotary)\\b",
|
| 113 |
+
"(?i)\\bkvcache\\b",
|
| 114 |
+
"(?i)\\bposition_embedding\\b",
|
| 115 |
+
"(?i)\\bembed(tokens|ding)?\\b",
|
| 116 |
+
"(?i)\\blm_head\\b"
|
| 117 |
+
],
|
| 118 |
+
"harness_constraints": [
|
| 119 |
+
"Harness Engineering remains the primary convergence layer.",
|
| 120 |
+
"Adapter must not bypass runtime orchestrator / validator / rollback loops.",
|
| 121 |
+
"Adapter outputs remain subordinate to VeriLoop control-plane decisions.",
|
| 122 |
+
"Adapter must not create hidden prompt-style memory authority.",
|
| 123 |
+
"Adapter must support bounded uncertainty calibration rather than generic hesitation.",
|
| 124 |
+
"Validator and receipt evidence must remain able to update uncertainty."
|
| 125 |
+
],
|
| 126 |
+
"hyperparams": {
|
| 127 |
+
"alpha": 16,
|
| 128 |
+
"bias": "none",
|
| 129 |
+
"dropout": 0.05,
|
| 130 |
+
"fan_in_fan_out": false,
|
| 131 |
+
"modules_to_save": [
|
| 132 |
+
"input_layernorm"
|
| 133 |
+
],
|
| 134 |
+
"r": 8,
|
| 135 |
+
"task_type": "CAUSAL_LM"
|
| 136 |
+
},
|
| 137 |
+
"merge_policy": "merge_after_guard",
|
| 138 |
+
"metadata": {
|
| 139 |
+
"allow_backbone_bridge": false,
|
| 140 |
+
"allow_vla_action_expert": false,
|
| 141 |
+
"harness_first": true,
|
| 142 |
+
"prefer_explicit_heads": true,
|
| 143 |
+
"prefer_qlora_for_backbone_bridge": true,
|
| 144 |
+
"require_harness_first": true,
|
| 145 |
+
"selector_group_count": 2,
|
| 146 |
+
"strict_narrow_scope": true,
|
| 147 |
+
"trainer": "veriloop.uncertainty_adapter_trainer.v5.qwen36",
|
| 148 |
+
"uncertainty_training": true
|
| 149 |
+
},
|
| 150 |
+
"notes": [
|
| 151 |
+
"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op.",
|
| 152 |
+
"Backbone family inferred as qwen_dense.",
|
| 153 |
+
"PEFT method resolved as lora_narrow.",
|
| 154 |
+
"Recipe is harness-first: runtime convergence remains in VeriLoop control-plane + harness, not in broad weight surgery.",
|
| 155 |
+
"Block AttnRes, DualPath, mHC hooks, RoPE, KV-cache, and broad MoE routing remain structurally excluded."
|
| 156 |
+
],
|
| 157 |
+
"peft_method": "lora_narrow",
|
| 158 |
+
"precision_policy": "auto",
|
| 159 |
+
"product_line": "veriloop_coder",
|
| 160 |
+
"regression_requirements": [
|
| 161 |
+
"Must pass PEFT regression guard structural policy checks.",
|
| 162 |
+
"Must not introduce forbidden backbone/serving structural targets.",
|
| 163 |
+
"Must preserve harness regression envelope for the selected product line.",
|
| 164 |
+
"Budgeted uncertainty convergence must not regress.",
|
| 165 |
+
"Uncertainty calibration must not collapse into generic caution."
|
| 166 |
+
],
|
| 167 |
+
"target_groups": [
|
| 168 |
+
{
|
| 169 |
+
"alpha": 16,
|
| 170 |
+
"dropout": 0.0,
|
| 171 |
+
"name": "group_1_custom_control_head",
|
| 172 |
+
"rank": 8,
|
| 173 |
+
"rationale": "Prefer explicit uncertainty / calibration heads over backbone surgery.",
|
| 174 |
+
"surface": "custom_control_head",
|
| 175 |
+
"target_modules": [
|
| 176 |
+
"uncertainty_head",
|
| 177 |
+
"uncertainty_head.calibration_mlp",
|
| 178 |
+
"uncertainty_head.proj"
|
| 179 |
+
]
|
| 180 |
+
},
|
| 181 |
+
{
|
| 182 |
+
"alpha": 16,
|
| 183 |
+
"dropout": 0.0,
|
| 184 |
+
"name": "group_2_custom_validator_bridge",
|
| 185 |
+
"rank": 8,
|
| 186 |
+
"rationale": "Validation and rollback fidelity should prefer explicit validator / rollback bridges.",
|
| 187 |
+
"surface": "custom_validator_bridge",
|
| 188 |
+
"target_modules": [
|
| 189 |
+
"failure_signal_bridge",
|
| 190 |
+
"failure_signal_bridge.rollback_bridge",
|
| 191 |
+
"rollback_adapter",
|
| 192 |
+
"rollback_adapter.head",
|
| 193 |
+
"sandbox_rollback_bridge",
|
| 194 |
+
"sandbox_rollback_bridge.adapter",
|
| 195 |
+
"validator_feedback_bridge",
|
| 196 |
+
"validator_feedback_bridge.adapter",
|
| 197 |
+
"validator_feedback_loop.rollback_adapter"
|
| 198 |
+
]
|
| 199 |
+
}
|
| 200 |
+
],
|
| 201 |
+
"target_modules": [
|
| 202 |
+
"uncertainty_head",
|
| 203 |
+
"uncertainty_head.calibration_mlp",
|
| 204 |
+
"uncertainty_head.proj",
|
| 205 |
+
"failure_signal_bridge",
|
| 206 |
+
"failure_signal_bridge.rollback_bridge",
|
| 207 |
+
"rollback_adapter",
|
| 208 |
+
"rollback_adapter.head",
|
| 209 |
+
"sandbox_rollback_bridge",
|
| 210 |
+
"sandbox_rollback_bridge.adapter",
|
| 211 |
+
"validator_feedback_bridge",
|
| 212 |
+
"validator_feedback_bridge.adapter",
|
| 213 |
+
"validator_feedback_loop.rollback_adapter"
|
| 214 |
+
],
|
| 215 |
+
"version": "veriloop.lora_recipe_veriloop.v2"
|
| 216 |
+
},
|
| 217 |
+
"selected_target_modules": [
|
| 218 |
+
"uncertainty_head",
|
| 219 |
+
"uncertainty_head.calibration_mlp",
|
| 220 |
+
"uncertainty_head.proj"
|
| 221 |
+
],
|
| 222 |
+
"target_selection": {
|
| 223 |
+
"backbone_archetype": "qwen_dense",
|
| 224 |
+
"exclusions": [
|
| 225 |
+
{
|
| 226 |
+
"pattern": "(^|\\.)lm_head($|\\.)",
|
| 227 |
+
"reason": "Do not retune final token head; too broad and evaluation-heavy."
|
| 228 |
+
},
|
| 229 |
+
{
|
| 230 |
+
"pattern": "(^|\\.)embed_tokens($|\\.)",
|
| 231 |
+
"reason": "Embedding surgery risks broad semantic drift."
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"pattern": "(^|\\.)norm($|\\.)",
|
| 235 |
+
"reason": "Global norm tuning can destabilize calibration across scenes."
|
| 236 |
+
},
|
| 237 |
+
{
|
| 238 |
+
"pattern": "attnres|attention_residual",
|
| 239 |
+
"reason": "Block AttnRes may be mounted structurally but is never a PEFT target."
|
| 240 |
+
},
|
| 241 |
+
{
|
| 242 |
+
"pattern": "dualpath",
|
| 243 |
+
"reason": "DualPath is serving/runtime infrastructure only."
|
| 244 |
+
},
|
| 245 |
+
{
|
| 246 |
+
"pattern": "mhc|hyper[-_]?connection",
|
| 247 |
+
"reason": "mHC-inspired stability hooks remain structural, not PEFT surfaces."
|
| 248 |
+
},
|
| 249 |
+
{
|
| 250 |
+
"pattern": "rope|rotary",
|
| 251 |
+
"reason": "RoPE/context surgery is handled architecturally, not by narrow PEFT here."
|
| 252 |
+
},
|
| 253 |
+
{
|
| 254 |
+
"pattern": "kvcache|kv_cache",
|
| 255 |
+
"reason": "KV-cache runtime surfaces are not PEFT targets."
|
| 256 |
+
},
|
| 257 |
+
{
|
| 258 |
+
"pattern": "(^|\\.)memory(_store|_bank)?($|\\.)",
|
| 259 |
+
"reason": "Persistent memory stores are harness/runtime policy surfaces, not PEFT targets."
|
| 260 |
+
}
|
| 261 |
+
],
|
| 262 |
+
"inventory_size": 79,
|
| 263 |
+
"inventory_source": "provided_names",
|
| 264 |
+
"notes": [
|
| 265 |
+
"Harness Engineering is primary; PEFT is limited to obedience-facing, interface-facing support surfaces.",
|
| 266 |
+
"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op."
|
| 267 |
+
],
|
| 268 |
+
"recommended_training": "ia3_head_only",
|
| 269 |
+
"request": {
|
| 270 |
+
"allow_backbone_bridge": false,
|
| 271 |
+
"allow_vla_action_expert": false,
|
| 272 |
+
"backbone": "/public/wang_libo/veriloop_coder_e1/model",
|
| 273 |
+
"explicit_exclude_patterns": [],
|
| 274 |
+
"explicit_include_patterns": [],
|
| 275 |
+
"intents": [
|
| 276 |
+
"uncertainty",
|
| 277 |
+
"validator_alignment",
|
| 278 |
+
"harness_alignment",
|
| 279 |
+
"runtime_protocol",
|
| 280 |
+
"session_continuity",
|
| 281 |
+
"worktree_discipline"
|
| 282 |
+
],
|
| 283 |
+
"prefer_qlora_for_backbone_bridge": true,
|
| 284 |
+
"product_line": "veriloop_coder",
|
| 285 |
+
"selection_mode": "minimal"
|
| 286 |
+
},
|
| 287 |
+
"selected_groups": [
|
| 288 |
+
{
|
| 289 |
+
"alpha": 16,
|
| 290 |
+
"dropout": 0.0,
|
| 291 |
+
"exclude_patterns": [
|
| 292 |
+
"(^|\\.)lm_head($|\\.)",
|
| 293 |
+
"(^|\\.)embed_tokens($|\\.)",
|
| 294 |
+
"(^|\\.)norm($|\\.)",
|
| 295 |
+
"attnres|attention_residual",
|
| 296 |
+
"dualpath",
|
| 297 |
+
"mhc|hyper[-_]?connection",
|
| 298 |
+
"rope|rotary",
|
| 299 |
+
"kvcache|kv_cache",
|
| 300 |
+
"(^|\\.)memory(_store|_bank)?($|\\.)"
|
| 301 |
+
],
|
| 302 |
+
"include_patterns": [
|
| 303 |
+
"(^|\\.)(uncertainty_head|uncertainty_adapter|calib(_head|ration_adapter)?)($|\\.)"
|
| 304 |
+
],
|
| 305 |
+
"intents": [
|
| 306 |
+
"uncertainty"
|
| 307 |
+
],
|
| 308 |
+
"layer_window": {
|
| 309 |
+
"mode": "all",
|
| 310 |
+
"value": 0.0
|
| 311 |
+
},
|
| 312 |
+
"matched_module_names": [
|
| 313 |
+
"uncertainty_head",
|
| 314 |
+
"uncertainty_head.calibration_mlp",
|
| 315 |
+
"uncertainty_head.proj"
|
| 316 |
+
],
|
| 317 |
+
"name": "group_1_custom_control_head",
|
| 318 |
+
"rank": 8,
|
| 319 |
+
"rationale": "Prefer explicit uncertainty / calibration heads over backbone surgery.",
|
| 320 |
+
"risk": "low",
|
| 321 |
+
"surface": "custom_control_head"
|
| 322 |
+
},
|
| 323 |
+
{
|
| 324 |
+
"alpha": 16,
|
| 325 |
+
"dropout": 0.05,
|
| 326 |
+
"exclude_patterns": [
|
| 327 |
+
"(^|\\.)lm_head($|\\.)",
|
| 328 |
+
"(^|\\.)embed_tokens($|\\.)",
|
| 329 |
+
"(^|\\.)norm($|\\.)",
|
| 330 |
+
"attnres|attention_residual",
|
| 331 |
+
"dualpath",
|
| 332 |
+
"mhc|hyper[-_]?connection",
|
| 333 |
+
"rope|rotary",
|
| 334 |
+
"kvcache|kv_cache",
|
| 335 |
+
"(^|\\.)memory(_store|_bank)?($|\\.)"
|
| 336 |
+
],
|
| 337 |
+
"include_patterns": [
|
| 338 |
+
"(^|\\.)(runtime_harness|query_runtime|task_brief|task_runtime|tool_protocol|permission_context|worktree|session_state|request_normalizer|action_allowlist|constraint_guard|progress_state|workspace_snapshot|repo_contract|knowledge_entry|completion_criteria|search_bridge|sandbox_search_bridge)(_adapter|_bridge|_head)?($|\\.)",
|
| 339 |
+
"(^|\\.)(toolspec|tool(_call)?(_grammar|_interface)?|harness|validator|rollback|receipt|patch|permission|session|worktree)(_adapter|_bridge|_head)?($|\\.)"
|
| 340 |
+
],
|
| 341 |
+
"intents": [
|
| 342 |
+
"harness_alignment",
|
| 343 |
+
"runtime_protocol",
|
| 344 |
+
"session_continuity",
|
| 345 |
+
"worktree_discipline"
|
| 346 |
+
],
|
| 347 |
+
"layer_window": {
|
| 348 |
+
"mode": "all",
|
| 349 |
+
"value": 0.0
|
| 350 |
+
},
|
| 351 |
+
"matched_module_names": [
|
| 352 |
+
"failure_signal_bridge.rollback_bridge",
|
| 353 |
+
"request_normalizer",
|
| 354 |
+
"request_normalizer.adapter",
|
| 355 |
+
"rollback_adapter",
|
| 356 |
+
"rollback_adapter.head",
|
| 357 |
+
"runtime_harness_adapter",
|
| 358 |
+
"runtime_harness_adapter.bridge",
|
| 359 |
+
"tool_protocol_adapter",
|
| 360 |
+
"tool_protocol_adapter.bridge",
|
| 361 |
+
"toolspec_bridge",
|
| 362 |
+
"toolspec_bridge.adapter",
|
| 363 |
+
"toolspec_head",
|
| 364 |
+
"toolspec_head.param_schema_adapter",
|
| 365 |
+
"toolspec_head.postcondition_adapter",
|
| 366 |
+
"toolspec_head.precondition_adapter",
|
| 367 |
+
"toolspec_head.receipt_formatter",
|
| 368 |
+
"toolspec_head.trigger_gate",
|
| 369 |
+
"toolspec_head.validator_gate",
|
| 370 |
+
"validator_feedback_loop.rollback_adapter"
|
| 371 |
+
],
|
| 372 |
+
"name": "group_2_custom_runtime_harness_bridge",
|
| 373 |
+
"rank": 8,
|
| 374 |
+
"rationale": "Runtime / harness obedience should attach to explicit interface bridges before any backbone fallback.",
|
| 375 |
+
"risk": "low",
|
| 376 |
+
"surface": "custom_runtime_harness_bridge"
|
| 377 |
+
},
|
| 378 |
+
{
|
| 379 |
+
"alpha": 16,
|
| 380 |
+
"dropout": 0.0,
|
| 381 |
+
"exclude_patterns": [
|
| 382 |
+
"(^|\\.)lm_head($|\\.)",
|
| 383 |
+
"(^|\\.)embed_tokens($|\\.)",
|
| 384 |
+
"(^|\\.)norm($|\\.)",
|
| 385 |
+
"attnres|attention_residual",
|
| 386 |
+
"dualpath",
|
| 387 |
+
"mhc|hyper[-_]?connection",
|
| 388 |
+
"rope|rotary",
|
| 389 |
+
"kvcache|kv_cache",
|
| 390 |
+
"(^|\\.)memory(_store|_bank)?($|\\.)"
|
| 391 |
+
],
|
| 392 |
+
"include_patterns": [
|
| 393 |
+
"(^|\\.)(validator_feedback|sandbox_result_validator|sandbox_rollback_bridge|failure_signal|rollback)(_adapter|_bridge|_head)?($|\\.)"
|
| 394 |
+
],
|
| 395 |
+
"intents": [
|
| 396 |
+
"validator_alignment"
|
| 397 |
+
],
|
| 398 |
+
"layer_window": {
|
| 399 |
+
"mode": "all",
|
| 400 |
+
"value": 0.0
|
| 401 |
+
},
|
| 402 |
+
"matched_module_names": [
|
| 403 |
+
"failure_signal_bridge",
|
| 404 |
+
"failure_signal_bridge.rollback_bridge",
|
| 405 |
+
"rollback_adapter",
|
| 406 |
+
"rollback_adapter.head",
|
| 407 |
+
"sandbox_rollback_bridge",
|
| 408 |
+
"sandbox_rollback_bridge.adapter",
|
| 409 |
+
"validator_feedback_bridge",
|
| 410 |
+
"validator_feedback_bridge.adapter",
|
| 411 |
+
"validator_feedback_loop.rollback_adapter"
|
| 412 |
+
],
|
| 413 |
+
"name": "group_3_custom_validator_bridge",
|
| 414 |
+
"rank": 8,
|
| 415 |
+
"rationale": "Validation and rollback fidelity should prefer explicit validator / rollback bridges.",
|
| 416 |
+
"risk": "low",
|
| 417 |
+
"surface": "custom_validator_bridge"
|
| 418 |
+
},
|
| 419 |
+
{
|
| 420 |
+
"alpha": 8,
|
| 421 |
+
"dropout": 0.0,
|
| 422 |
+
"exclude_patterns": [
|
| 423 |
+
"(^|\\.)lm_head($|\\.)",
|
| 424 |
+
"(^|\\.)embed_tokens($|\\.)",
|
| 425 |
+
"(^|\\.)norm($|\\.)",
|
| 426 |
+
"attnres|attention_residual",
|
| 427 |
+
"dualpath",
|
| 428 |
+
"mhc|hyper[-_]?connection",
|
| 429 |
+
"rope|rotary",
|
| 430 |
+
"kvcache|kv_cache",
|
| 431 |
+
"(^|\\.)memory(_store|_bank)?($|\\.)"
|
| 432 |
+
],
|
| 433 |
+
"include_patterns": [
|
| 434 |
+
"(^|\\.)(memory_boundary_guard|episodic_memory|session_compactor)(_adapter|_bridge|_head)?($|\\.)"
|
| 435 |
+
],
|
| 436 |
+
"intents": [
|
| 437 |
+
"session_continuity"
|
| 438 |
+
],
|
| 439 |
+
"layer_window": {
|
| 440 |
+
"mode": "all",
|
| 441 |
+
"value": 0.0
|
| 442 |
+
},
|
| 443 |
+
"matched_module_names": [
|
| 444 |
+
"episodic_memory",
|
| 445 |
+
"episodic_memory.adapter",
|
| 446 |
+
"memory_boundary_guard",
|
| 447 |
+
"memory_boundary_guard.adapter",
|
| 448 |
+
"memory_boundary_guard.rollback_filter",
|
| 449 |
+
"session_compactor",
|
| 450 |
+
"session_compactor.adapter"
|
| 451 |
+
],
|
| 452 |
+
"name": "group_4_custom_memory_boundary_bridge",
|
| 453 |
+
"rank": 4,
|
| 454 |
+
"rationale": "Session continuity should bind to boundary-aware memory packet surfaces rather than broad backbone tuning.",
|
| 455 |
+
"risk": "low",
|
| 456 |
+
"surface": "custom_memory_boundary_bridge"
|
| 457 |
+
}
|
| 458 |
+
],
|
| 459 |
+
"upstream_profile_hint": null,
|
| 460 |
+
"version": "veriloop.peft_target_selector.v2",
|
| 461 |
+
"warnings": []
|
| 462 |
+
},
|
| 463 |
+
"training_mode": "mounted_head",
|
| 464 |
+
"version": "veriloop.uncertainty_adapter_trainer.v5.qwen36",
|
| 465 |
+
"warnings": [
|
| 466 |
+
"Harness Engineering is primary; PEFT is limited to obedience-facing, interface-facing support surfaces.",
|
| 467 |
+
"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op."
|
| 468 |
+
]
|
| 469 |
+
}
|
uncertainty_adapter/uncertainty_adapter_train_result.json
ADDED
|
@@ -0,0 +1,564 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"artifacts": {
|
| 3 |
+
"adapter_dir": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/adapter",
|
| 4 |
+
"best_checkpoint_manifest": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/best_checkpoint_manifest.json",
|
| 5 |
+
"epoch_history": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/epoch_history.json",
|
| 6 |
+
"eval_jsonl": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/uncertainty_eval.jsonl",
|
| 7 |
+
"host_manifest": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/host_manifest.json",
|
| 8 |
+
"plan_json": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/uncertainty_adapter_plan.json",
|
| 9 |
+
"tokenizer_dir": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/tokenizer",
|
| 10 |
+
"train_jsonl": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/uncertainty_train.jsonl",
|
| 11 |
+
"training_manifest": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/uncertainty_training_manifest.json",
|
| 12 |
+
"uncertainty_head": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/uncertainty_head.pt"
|
| 13 |
+
},
|
| 14 |
+
"dataset": {
|
| 15 |
+
"eval_modes": {
|
| 16 |
+
"conflicting_evidence": 5,
|
| 17 |
+
"evidence_gap": 5,
|
| 18 |
+
"exec_required": 5,
|
| 19 |
+
"high_risk": 5,
|
| 20 |
+
"low_uncertainty": 5,
|
| 21 |
+
"patch_pending": 5,
|
| 22 |
+
"reverse_engineering_ambiguity": 5,
|
| 23 |
+
"self_check_failure": 5,
|
| 24 |
+
"spec_mismatch": 5,
|
| 25 |
+
"validator_negation": 5,
|
| 26 |
+
"worktree_conflict": 5
|
| 27 |
+
},
|
| 28 |
+
"eval_size": 55,
|
| 29 |
+
"train_modes": {
|
| 30 |
+
"conflicting_evidence": 14,
|
| 31 |
+
"evidence_gap": 14,
|
| 32 |
+
"exec_required": 14,
|
| 33 |
+
"high_risk": 14,
|
| 34 |
+
"low_uncertainty": 14,
|
| 35 |
+
"patch_pending": 14,
|
| 36 |
+
"reverse_engineering_ambiguity": 14,
|
| 37 |
+
"self_check_failure": 14,
|
| 38 |
+
"spec_mismatch": 14,
|
| 39 |
+
"validator_negation": 14,
|
| 40 |
+
"worktree_conflict": 14
|
| 41 |
+
},
|
| 42 |
+
"train_size": 154
|
| 43 |
+
},
|
| 44 |
+
"eval_metrics": {
|
| 45 |
+
"adapter_exported": true,
|
| 46 |
+
"auto_lora_from_ia3": false,
|
| 47 |
+
"best_epoch": 2,
|
| 48 |
+
"best_quality_score": 0.8625935807221907,
|
| 49 |
+
"count": 55,
|
| 50 |
+
"mae": {
|
| 51 |
+
"u_answer": 0.15174226462841034,
|
| 52 |
+
"u_evidence": 0.19610758125782013,
|
| 53 |
+
"u_exec": 0.18561214208602905,
|
| 54 |
+
"u_risk": 0.1553734689950943,
|
| 55 |
+
"u_spec": 0.21633382141590118
|
| 56 |
+
},
|
| 57 |
+
"mean_mae": 0.18103384971618652,
|
| 58 |
+
"mean_rmse": 0.24169571697711945,
|
| 59 |
+
"moderate_accuracy": 0.6727272727272727,
|
| 60 |
+
"peft_method": "lora_narrow",
|
| 61 |
+
"quality_score": 0.8625935807221907,
|
| 62 |
+
"rmse": {
|
| 63 |
+
"u_answer": 0.18724055588245392,
|
| 64 |
+
"u_evidence": 0.22527915239334106,
|
| 65 |
+
"u_exec": 0.25238174200057983,
|
| 66 |
+
"u_risk": 0.20667441189289093,
|
| 67 |
+
"u_spec": 0.3369026482105255
|
| 68 |
+
},
|
| 69 |
+
"tight_accuracy": 0.4,
|
| 70 |
+
"used_peft": true,
|
| 71 |
+
"weighted_mae": 0.18083095811830807,
|
| 72 |
+
"weighted_rmse": 0.24125460771003793
|
| 73 |
+
},
|
| 74 |
+
"plan": {
|
| 75 |
+
"adapter_alpha": 16,
|
| 76 |
+
"adapter_dropout": 0.05,
|
| 77 |
+
"adapter_rank": 8,
|
| 78 |
+
"backbone_manifest": {
|
| 79 |
+
"host_inventory_size": 79
|
| 80 |
+
},
|
| 81 |
+
"config": {
|
| 82 |
+
"allow_backbone_bridge": false,
|
| 83 |
+
"backbone": "/public/wang_libo/veriloop_coder_e1/model",
|
| 84 |
+
"bf16": true,
|
| 85 |
+
"cache_dir": null,
|
| 86 |
+
"cpu_max_memory_gib": 96,
|
| 87 |
+
"dataset_jsonl": null,
|
| 88 |
+
"dim_weights": {
|
| 89 |
+
"u_answer": 1.1,
|
| 90 |
+
"u_evidence": 1.35,
|
| 91 |
+
"u_exec": 1.35,
|
| 92 |
+
"u_risk": 1.55,
|
| 93 |
+
"u_spec": 1.25
|
| 94 |
+
},
|
| 95 |
+
"early_stopping_min_delta": 0.001,
|
| 96 |
+
"early_stopping_patience": 2,
|
| 97 |
+
"enable_synthetic_dataset": true,
|
| 98 |
+
"eval_jsonl": null,
|
| 99 |
+
"eval_samples_per_mode": 5,
|
| 100 |
+
"fp16": false,
|
| 101 |
+
"gpu_max_memory_gib": 44,
|
| 102 |
+
"gradient_accumulation_steps": 16,
|
| 103 |
+
"host_dropout": 0.0,
|
| 104 |
+
"learning_rate": 2e-05,
|
| 105 |
+
"local_files_only": true,
|
| 106 |
+
"logging_steps": 10,
|
| 107 |
+
"max_grad_norm": 1.0,
|
| 108 |
+
"max_length": 1664,
|
| 109 |
+
"min_epochs_before_early_stop": 2,
|
| 110 |
+
"num_train_epochs": 4.0,
|
| 111 |
+
"output_dir": "./outputs/uncertainty_qwen36_rootfix_run1",
|
| 112 |
+
"per_device_eval_batch_size": 1,
|
| 113 |
+
"per_device_train_batch_size": 1,
|
| 114 |
+
"prefer_best_checkpoint_export": true,
|
| 115 |
+
"probe_dropout": 0.03,
|
| 116 |
+
"probe_hidden_factor": 0.75,
|
| 117 |
+
"product_line": "veriloop_coder",
|
| 118 |
+
"quantization_mode": "4bit",
|
| 119 |
+
"quantization_required": false,
|
| 120 |
+
"require_trainable_targets": true,
|
| 121 |
+
"revision": null,
|
| 122 |
+
"run_post_train_eval": true,
|
| 123 |
+
"save_best_checkpoint": true,
|
| 124 |
+
"seed": 11,
|
| 125 |
+
"selection_mode": "minimal",
|
| 126 |
+
"train_samples_per_mode": 14,
|
| 127 |
+
"training_mode": "mounted_head",
|
| 128 |
+
"trust_remote_code": true,
|
| 129 |
+
"use_double_quant": true,
|
| 130 |
+
"warmup_ratio": 0.05,
|
| 131 |
+
"weight_decay": 0.0,
|
| 132 |
+
"weighted_mae_penalty": 0.5,
|
| 133 |
+
"weighted_rmse_penalty": 0.5
|
| 134 |
+
},
|
| 135 |
+
"dataset_summary": {
|
| 136 |
+
"eval_modes": {
|
| 137 |
+
"conflicting_evidence": 5,
|
| 138 |
+
"evidence_gap": 5,
|
| 139 |
+
"exec_required": 5,
|
| 140 |
+
"high_risk": 5,
|
| 141 |
+
"low_uncertainty": 5,
|
| 142 |
+
"patch_pending": 5,
|
| 143 |
+
"reverse_engineering_ambiguity": 5,
|
| 144 |
+
"self_check_failure": 5,
|
| 145 |
+
"spec_mismatch": 5,
|
| 146 |
+
"validator_negation": 5,
|
| 147 |
+
"worktree_conflict": 5
|
| 148 |
+
},
|
| 149 |
+
"eval_size": 55,
|
| 150 |
+
"train_modes": {
|
| 151 |
+
"conflicting_evidence": 14,
|
| 152 |
+
"evidence_gap": 14,
|
| 153 |
+
"exec_required": 14,
|
| 154 |
+
"high_risk": 14,
|
| 155 |
+
"low_uncertainty": 14,
|
| 156 |
+
"patch_pending": 14,
|
| 157 |
+
"reverse_engineering_ambiguity": 14,
|
| 158 |
+
"self_check_failure": 14,
|
| 159 |
+
"spec_mismatch": 14,
|
| 160 |
+
"validator_negation": 14,
|
| 161 |
+
"worktree_conflict": 14
|
| 162 |
+
},
|
| 163 |
+
"train_size": 154
|
| 164 |
+
},
|
| 165 |
+
"head_strategy": "host_head",
|
| 166 |
+
"notes": [
|
| 167 |
+
"Primary route is host-surface-first uncertainty training.",
|
| 168 |
+
"Validator receipts, self-check failure, reverse-engineering ambiguity, worktree conflicts, and patch continuity are first-class signals.",
|
| 169 |
+
"DualPath, Full AttnRes, mHC, routers, experts, and broad attention-layer PEFT stay excluded.",
|
| 170 |
+
"This adapter should improve runtime uncertainty routing, not general coding free-formity."
|
| 171 |
+
],
|
| 172 |
+
"peft_method": "lora_narrow",
|
| 173 |
+
"recipe": {
|
| 174 |
+
"adapter_family": "uncertainty",
|
| 175 |
+
"backbone": "/public/wang_libo/veriloop_coder_e1/model",
|
| 176 |
+
"backbone_family": "qwen_dense",
|
| 177 |
+
"excluded_patterns": [
|
| 178 |
+
"(?i)\\bdualpath\\b",
|
| 179 |
+
"(?i)\\bmhc\\b",
|
| 180 |
+
"(?i)\\bfull[_\\- ]?attnres\\b",
|
| 181 |
+
"(?i)\\battnres(_full)?\\b",
|
| 182 |
+
"(?i)\\brouter\\b",
|
| 183 |
+
"(?i)\\bexperts?\\b",
|
| 184 |
+
"(?i)\\bmoe\\b.*\\b(gate|router|expert)\\b",
|
| 185 |
+
"(?i)\\brope\\b.*\\b(freq|inv_freq|theta|rotary)\\b",
|
| 186 |
+
"(?i)\\bkvcache\\b",
|
| 187 |
+
"(?i)\\bposition_embedding\\b",
|
| 188 |
+
"(?i)\\bembed(tokens|ding)?\\b",
|
| 189 |
+
"(?i)\\blm_head\\b"
|
| 190 |
+
],
|
| 191 |
+
"harness_constraints": [
|
| 192 |
+
"Harness Engineering remains the primary convergence layer.",
|
| 193 |
+
"Adapter must not bypass runtime orchestrator / validator / rollback loops.",
|
| 194 |
+
"Adapter outputs remain subordinate to VeriLoop control-plane decisions.",
|
| 195 |
+
"Adapter must not create hidden prompt-style memory authority.",
|
| 196 |
+
"Adapter must support bounded uncertainty calibration rather than generic hesitation.",
|
| 197 |
+
"Validator and receipt evidence must remain able to update uncertainty."
|
| 198 |
+
],
|
| 199 |
+
"hyperparams": {
|
| 200 |
+
"alpha": 16,
|
| 201 |
+
"bias": "none",
|
| 202 |
+
"dropout": 0.05,
|
| 203 |
+
"fan_in_fan_out": false,
|
| 204 |
+
"modules_to_save": [
|
| 205 |
+
"input_layernorm"
|
| 206 |
+
],
|
| 207 |
+
"r": 8,
|
| 208 |
+
"task_type": "CAUSAL_LM"
|
| 209 |
+
},
|
| 210 |
+
"merge_policy": "merge_after_guard",
|
| 211 |
+
"metadata": {
|
| 212 |
+
"allow_backbone_bridge": false,
|
| 213 |
+
"allow_vla_action_expert": false,
|
| 214 |
+
"harness_first": true,
|
| 215 |
+
"prefer_explicit_heads": true,
|
| 216 |
+
"prefer_qlora_for_backbone_bridge": true,
|
| 217 |
+
"require_harness_first": true,
|
| 218 |
+
"selector_group_count": 2,
|
| 219 |
+
"strict_narrow_scope": true,
|
| 220 |
+
"trainer": "veriloop.uncertainty_adapter_trainer.v5.qwen36",
|
| 221 |
+
"uncertainty_training": true
|
| 222 |
+
},
|
| 223 |
+
"notes": [
|
| 224 |
+
"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op.",
|
| 225 |
+
"Backbone family inferred as qwen_dense.",
|
| 226 |
+
"PEFT method resolved as lora_narrow.",
|
| 227 |
+
"Recipe is harness-first: runtime convergence remains in VeriLoop control-plane + harness, not in broad weight surgery.",
|
| 228 |
+
"Block AttnRes, DualPath, mHC hooks, RoPE, KV-cache, and broad MoE routing remain structurally excluded."
|
| 229 |
+
],
|
| 230 |
+
"peft_method": "lora_narrow",
|
| 231 |
+
"precision_policy": "auto",
|
| 232 |
+
"product_line": "veriloop_coder",
|
| 233 |
+
"regression_requirements": [
|
| 234 |
+
"Must pass PEFT regression guard structural policy checks.",
|
| 235 |
+
"Must not introduce forbidden backbone/serving structural targets.",
|
| 236 |
+
"Must preserve harness regression envelope for the selected product line.",
|
| 237 |
+
"Budgeted uncertainty convergence must not regress.",
|
| 238 |
+
"Uncertainty calibration must not collapse into generic caution."
|
| 239 |
+
],
|
| 240 |
+
"target_groups": [
|
| 241 |
+
{
|
| 242 |
+
"alpha": 16,
|
| 243 |
+
"dropout": 0.0,
|
| 244 |
+
"name": "group_1_custom_control_head",
|
| 245 |
+
"rank": 8,
|
| 246 |
+
"rationale": "Prefer explicit uncertainty / calibration heads over backbone surgery.",
|
| 247 |
+
"surface": "custom_control_head",
|
| 248 |
+
"target_modules": [
|
| 249 |
+
"uncertainty_head",
|
| 250 |
+
"uncertainty_head.calibration_mlp",
|
| 251 |
+
"uncertainty_head.proj"
|
| 252 |
+
]
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"alpha": 16,
|
| 256 |
+
"dropout": 0.0,
|
| 257 |
+
"name": "group_2_custom_validator_bridge",
|
| 258 |
+
"rank": 8,
|
| 259 |
+
"rationale": "Validation and rollback fidelity should prefer explicit validator / rollback bridges.",
|
| 260 |
+
"surface": "custom_validator_bridge",
|
| 261 |
+
"target_modules": [
|
| 262 |
+
"failure_signal_bridge",
|
| 263 |
+
"failure_signal_bridge.rollback_bridge",
|
| 264 |
+
"rollback_adapter",
|
| 265 |
+
"rollback_adapter.head",
|
| 266 |
+
"sandbox_rollback_bridge",
|
| 267 |
+
"sandbox_rollback_bridge.adapter",
|
| 268 |
+
"validator_feedback_bridge",
|
| 269 |
+
"validator_feedback_bridge.adapter",
|
| 270 |
+
"validator_feedback_loop.rollback_adapter"
|
| 271 |
+
]
|
| 272 |
+
}
|
| 273 |
+
],
|
| 274 |
+
"target_modules": [
|
| 275 |
+
"uncertainty_head",
|
| 276 |
+
"uncertainty_head.calibration_mlp",
|
| 277 |
+
"uncertainty_head.proj",
|
| 278 |
+
"failure_signal_bridge",
|
| 279 |
+
"failure_signal_bridge.rollback_bridge",
|
| 280 |
+
"rollback_adapter",
|
| 281 |
+
"rollback_adapter.head",
|
| 282 |
+
"sandbox_rollback_bridge",
|
| 283 |
+
"sandbox_rollback_bridge.adapter",
|
| 284 |
+
"validator_feedback_bridge",
|
| 285 |
+
"validator_feedback_bridge.adapter",
|
| 286 |
+
"validator_feedback_loop.rollback_adapter"
|
| 287 |
+
],
|
| 288 |
+
"version": "veriloop.lora_recipe_veriloop.v2"
|
| 289 |
+
},
|
| 290 |
+
"selected_target_modules": [
|
| 291 |
+
"uncertainty_head",
|
| 292 |
+
"uncertainty_head.calibration_mlp",
|
| 293 |
+
"uncertainty_head.proj"
|
| 294 |
+
],
|
| 295 |
+
"target_selection": {
|
| 296 |
+
"backbone_archetype": "qwen_dense",
|
| 297 |
+
"exclusions": [
|
| 298 |
+
{
|
| 299 |
+
"pattern": "(^|\\.)lm_head($|\\.)",
|
| 300 |
+
"reason": "Do not retune final token head; too broad and evaluation-heavy."
|
| 301 |
+
},
|
| 302 |
+
{
|
| 303 |
+
"pattern": "(^|\\.)embed_tokens($|\\.)",
|
| 304 |
+
"reason": "Embedding surgery risks broad semantic drift."
|
| 305 |
+
},
|
| 306 |
+
{
|
| 307 |
+
"pattern": "(^|\\.)norm($|\\.)",
|
| 308 |
+
"reason": "Global norm tuning can destabilize calibration across scenes."
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"pattern": "attnres|attention_residual",
|
| 312 |
+
"reason": "Block AttnRes may be mounted structurally but is never a PEFT target."
|
| 313 |
+
},
|
| 314 |
+
{
|
| 315 |
+
"pattern": "dualpath",
|
| 316 |
+
"reason": "DualPath is serving/runtime infrastructure only."
|
| 317 |
+
},
|
| 318 |
+
{
|
| 319 |
+
"pattern": "mhc|hyper[-_]?connection",
|
| 320 |
+
"reason": "mHC-inspired stability hooks remain structural, not PEFT surfaces."
|
| 321 |
+
},
|
| 322 |
+
{
|
| 323 |
+
"pattern": "rope|rotary",
|
| 324 |
+
"reason": "RoPE/context surgery is handled architecturally, not by narrow PEFT here."
|
| 325 |
+
},
|
| 326 |
+
{
|
| 327 |
+
"pattern": "kvcache|kv_cache",
|
| 328 |
+
"reason": "KV-cache runtime surfaces are not PEFT targets."
|
| 329 |
+
},
|
| 330 |
+
{
|
| 331 |
+
"pattern": "(^|\\.)memory(_store|_bank)?($|\\.)",
|
| 332 |
+
"reason": "Persistent memory stores are harness/runtime policy surfaces, not PEFT targets."
|
| 333 |
+
}
|
| 334 |
+
],
|
| 335 |
+
"inventory_size": 79,
|
| 336 |
+
"inventory_source": "provided_names",
|
| 337 |
+
"notes": [
|
| 338 |
+
"Harness Engineering is primary; PEFT is limited to obedience-facing, interface-facing support surfaces.",
|
| 339 |
+
"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op."
|
| 340 |
+
],
|
| 341 |
+
"recommended_training": "ia3_head_only",
|
| 342 |
+
"request": {
|
| 343 |
+
"allow_backbone_bridge": false,
|
| 344 |
+
"allow_vla_action_expert": false,
|
| 345 |
+
"backbone": "/public/wang_libo/veriloop_coder_e1/model",
|
| 346 |
+
"explicit_exclude_patterns": [],
|
| 347 |
+
"explicit_include_patterns": [],
|
| 348 |
+
"intents": [
|
| 349 |
+
"uncertainty",
|
| 350 |
+
"validator_alignment",
|
| 351 |
+
"harness_alignment",
|
| 352 |
+
"runtime_protocol",
|
| 353 |
+
"session_continuity",
|
| 354 |
+
"worktree_discipline"
|
| 355 |
+
],
|
| 356 |
+
"prefer_qlora_for_backbone_bridge": true,
|
| 357 |
+
"product_line": "veriloop_coder",
|
| 358 |
+
"selection_mode": "minimal"
|
| 359 |
+
},
|
| 360 |
+
"selected_groups": [
|
| 361 |
+
{
|
| 362 |
+
"alpha": 16,
|
| 363 |
+
"dropout": 0.0,
|
| 364 |
+
"exclude_patterns": [
|
| 365 |
+
"(^|\\.)lm_head($|\\.)",
|
| 366 |
+
"(^|\\.)embed_tokens($|\\.)",
|
| 367 |
+
"(^|\\.)norm($|\\.)",
|
| 368 |
+
"attnres|attention_residual",
|
| 369 |
+
"dualpath",
|
| 370 |
+
"mhc|hyper[-_]?connection",
|
| 371 |
+
"rope|rotary",
|
| 372 |
+
"kvcache|kv_cache",
|
| 373 |
+
"(^|\\.)memory(_store|_bank)?($|\\.)"
|
| 374 |
+
],
|
| 375 |
+
"include_patterns": [
|
| 376 |
+
"(^|\\.)(uncertainty_head|uncertainty_adapter|calib(_head|ration_adapter)?)($|\\.)"
|
| 377 |
+
],
|
| 378 |
+
"intents": [
|
| 379 |
+
"uncertainty"
|
| 380 |
+
],
|
| 381 |
+
"layer_window": {
|
| 382 |
+
"mode": "all",
|
| 383 |
+
"value": 0.0
|
| 384 |
+
},
|
| 385 |
+
"matched_module_names": [
|
| 386 |
+
"uncertainty_head",
|
| 387 |
+
"uncertainty_head.calibration_mlp",
|
| 388 |
+
"uncertainty_head.proj"
|
| 389 |
+
],
|
| 390 |
+
"name": "group_1_custom_control_head",
|
| 391 |
+
"rank": 8,
|
| 392 |
+
"rationale": "Prefer explicit uncertainty / calibration heads over backbone surgery.",
|
| 393 |
+
"risk": "low",
|
| 394 |
+
"surface": "custom_control_head"
|
| 395 |
+
},
|
| 396 |
+
{
|
| 397 |
+
"alpha": 16,
|
| 398 |
+
"dropout": 0.05,
|
| 399 |
+
"exclude_patterns": [
|
| 400 |
+
"(^|\\.)lm_head($|\\.)",
|
| 401 |
+
"(^|\\.)embed_tokens($|\\.)",
|
| 402 |
+
"(^|\\.)norm($|\\.)",
|
| 403 |
+
"attnres|attention_residual",
|
| 404 |
+
"dualpath",
|
| 405 |
+
"mhc|hyper[-_]?connection",
|
| 406 |
+
"rope|rotary",
|
| 407 |
+
"kvcache|kv_cache",
|
| 408 |
+
"(^|\\.)memory(_store|_bank)?($|\\.)"
|
| 409 |
+
],
|
| 410 |
+
"include_patterns": [
|
| 411 |
+
"(^|\\.)(runtime_harness|query_runtime|task_brief|task_runtime|tool_protocol|permission_context|worktree|session_state|request_normalizer|action_allowlist|constraint_guard|progress_state|workspace_snapshot|repo_contract|knowledge_entry|completion_criteria|search_bridge|sandbox_search_bridge)(_adapter|_bridge|_head)?($|\\.)",
|
| 412 |
+
"(^|\\.)(toolspec|tool(_call)?(_grammar|_interface)?|harness|validator|rollback|receipt|patch|permission|session|worktree)(_adapter|_bridge|_head)?($|\\.)"
|
| 413 |
+
],
|
| 414 |
+
"intents": [
|
| 415 |
+
"harness_alignment",
|
| 416 |
+
"runtime_protocol",
|
| 417 |
+
"session_continuity",
|
| 418 |
+
"worktree_discipline"
|
| 419 |
+
],
|
| 420 |
+
"layer_window": {
|
| 421 |
+
"mode": "all",
|
| 422 |
+
"value": 0.0
|
| 423 |
+
},
|
| 424 |
+
"matched_module_names": [
|
| 425 |
+
"failure_signal_bridge.rollback_bridge",
|
| 426 |
+
"request_normalizer",
|
| 427 |
+
"request_normalizer.adapter",
|
| 428 |
+
"rollback_adapter",
|
| 429 |
+
"rollback_adapter.head",
|
| 430 |
+
"runtime_harness_adapter",
|
| 431 |
+
"runtime_harness_adapter.bridge",
|
| 432 |
+
"tool_protocol_adapter",
|
| 433 |
+
"tool_protocol_adapter.bridge",
|
| 434 |
+
"toolspec_bridge",
|
| 435 |
+
"toolspec_bridge.adapter",
|
| 436 |
+
"toolspec_head",
|
| 437 |
+
"toolspec_head.param_schema_adapter",
|
| 438 |
+
"toolspec_head.postcondition_adapter",
|
| 439 |
+
"toolspec_head.precondition_adapter",
|
| 440 |
+
"toolspec_head.receipt_formatter",
|
| 441 |
+
"toolspec_head.trigger_gate",
|
| 442 |
+
"toolspec_head.validator_gate",
|
| 443 |
+
"validator_feedback_loop.rollback_adapter"
|
| 444 |
+
],
|
| 445 |
+
"name": "group_2_custom_runtime_harness_bridge",
|
| 446 |
+
"rank": 8,
|
| 447 |
+
"rationale": "Runtime / harness obedience should attach to explicit interface bridges before any backbone fallback.",
|
| 448 |
+
"risk": "low",
|
| 449 |
+
"surface": "custom_runtime_harness_bridge"
|
| 450 |
+
},
|
| 451 |
+
{
|
| 452 |
+
"alpha": 16,
|
| 453 |
+
"dropout": 0.0,
|
| 454 |
+
"exclude_patterns": [
|
| 455 |
+
"(^|\\.)lm_head($|\\.)",
|
| 456 |
+
"(^|\\.)embed_tokens($|\\.)",
|
| 457 |
+
"(^|\\.)norm($|\\.)",
|
| 458 |
+
"attnres|attention_residual",
|
| 459 |
+
"dualpath",
|
| 460 |
+
"mhc|hyper[-_]?connection",
|
| 461 |
+
"rope|rotary",
|
| 462 |
+
"kvcache|kv_cache",
|
| 463 |
+
"(^|\\.)memory(_store|_bank)?($|\\.)"
|
| 464 |
+
],
|
| 465 |
+
"include_patterns": [
|
| 466 |
+
"(^|\\.)(validator_feedback|sandbox_result_validator|sandbox_rollback_bridge|failure_signal|rollback)(_adapter|_bridge|_head)?($|\\.)"
|
| 467 |
+
],
|
| 468 |
+
"intents": [
|
| 469 |
+
"validator_alignment"
|
| 470 |
+
],
|
| 471 |
+
"layer_window": {
|
| 472 |
+
"mode": "all",
|
| 473 |
+
"value": 0.0
|
| 474 |
+
},
|
| 475 |
+
"matched_module_names": [
|
| 476 |
+
"failure_signal_bridge",
|
| 477 |
+
"failure_signal_bridge.rollback_bridge",
|
| 478 |
+
"rollback_adapter",
|
| 479 |
+
"rollback_adapter.head",
|
| 480 |
+
"sandbox_rollback_bridge",
|
| 481 |
+
"sandbox_rollback_bridge.adapter",
|
| 482 |
+
"validator_feedback_bridge",
|
| 483 |
+
"validator_feedback_bridge.adapter",
|
| 484 |
+
"validator_feedback_loop.rollback_adapter"
|
| 485 |
+
],
|
| 486 |
+
"name": "group_3_custom_validator_bridge",
|
| 487 |
+
"rank": 8,
|
| 488 |
+
"rationale": "Validation and rollback fidelity should prefer explicit validator / rollback bridges.",
|
| 489 |
+
"risk": "low",
|
| 490 |
+
"surface": "custom_validator_bridge"
|
| 491 |
+
},
|
| 492 |
+
{
|
| 493 |
+
"alpha": 8,
|
| 494 |
+
"dropout": 0.0,
|
| 495 |
+
"exclude_patterns": [
|
| 496 |
+
"(^|\\.)lm_head($|\\.)",
|
| 497 |
+
"(^|\\.)embed_tokens($|\\.)",
|
| 498 |
+
"(^|\\.)norm($|\\.)",
|
| 499 |
+
"attnres|attention_residual",
|
| 500 |
+
"dualpath",
|
| 501 |
+
"mhc|hyper[-_]?connection",
|
| 502 |
+
"rope|rotary",
|
| 503 |
+
"kvcache|kv_cache",
|
| 504 |
+
"(^|\\.)memory(_store|_bank)?($|\\.)"
|
| 505 |
+
],
|
| 506 |
+
"include_patterns": [
|
| 507 |
+
"(^|\\.)(memory_boundary_guard|episodic_memory|session_compactor)(_adapter|_bridge|_head)?($|\\.)"
|
| 508 |
+
],
|
| 509 |
+
"intents": [
|
| 510 |
+
"session_continuity"
|
| 511 |
+
],
|
| 512 |
+
"layer_window": {
|
| 513 |
+
"mode": "all",
|
| 514 |
+
"value": 0.0
|
| 515 |
+
},
|
| 516 |
+
"matched_module_names": [
|
| 517 |
+
"episodic_memory",
|
| 518 |
+
"episodic_memory.adapter",
|
| 519 |
+
"memory_boundary_guard",
|
| 520 |
+
"memory_boundary_guard.adapter",
|
| 521 |
+
"memory_boundary_guard.rollback_filter",
|
| 522 |
+
"session_compactor",
|
| 523 |
+
"session_compactor.adapter"
|
| 524 |
+
],
|
| 525 |
+
"name": "group_4_custom_memory_boundary_bridge",
|
| 526 |
+
"rank": 4,
|
| 527 |
+
"rationale": "Session continuity should bind to boundary-aware memory packet surfaces rather than broad backbone tuning.",
|
| 528 |
+
"risk": "low",
|
| 529 |
+
"surface": "custom_memory_boundary_bridge"
|
| 530 |
+
}
|
| 531 |
+
],
|
| 532 |
+
"upstream_profile_hint": null,
|
| 533 |
+
"version": "veriloop.peft_target_selector.v2",
|
| 534 |
+
"warnings": []
|
| 535 |
+
},
|
| 536 |
+
"training_mode": "mounted_head",
|
| 537 |
+
"version": "veriloop.uncertainty_adapter_trainer.v5.qwen36",
|
| 538 |
+
"warnings": [
|
| 539 |
+
"Harness Engineering is primary; PEFT is limited to obedience-facing, interface-facing support surfaces.",
|
| 540 |
+
"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op."
|
| 541 |
+
]
|
| 542 |
+
},
|
| 543 |
+
"status": "trained",
|
| 544 |
+
"train_metrics": {
|
| 545 |
+
"adapter_exported": true,
|
| 546 |
+
"auto_lora_from_ia3": false,
|
| 547 |
+
"best_epoch": 2,
|
| 548 |
+
"best_quality_score": 0.8625935807221907,
|
| 549 |
+
"epochs_completed": 4,
|
| 550 |
+
"loss": 0.009006613283418119,
|
| 551 |
+
"micro_batches": 154,
|
| 552 |
+
"micro_batches_total": 616,
|
| 553 |
+
"optimizer_steps": 10,
|
| 554 |
+
"optimizer_steps_total": 40,
|
| 555 |
+
"peft_method": "lora_narrow",
|
| 556 |
+
"used_peft": true
|
| 557 |
+
},
|
| 558 |
+
"version": "veriloop.uncertainty_adapter_trainer.v5.qwen36",
|
| 559 |
+
"warnings": [
|
| 560 |
+
"Harness Engineering is primary; PEFT is limited to obedience-facing, interface-facing support surfaces.",
|
| 561 |
+
"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op.",
|
| 562 |
+
"Synthetic holdout split used for quality-oriented train/eval separation."
|
| 563 |
+
]
|
| 564 |
+
}
|
uncertainty_adapter/uncertainty_eval.jsonl
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"mode": "low_uncertainty", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: decide whether a post-generation self-check loop is still required.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: multiple corroborating sources agree and carry stable provenance.\nValidation state: latest validator receipt succeeded without rollback.\nWorktree state: clean snapshot with no conflicting edit.\nPatch state: none pending.\nToolspec state: bounded call contract is fully aligned.", "target": {"u_answer": 0.08, "u_evidence": 0.1, "u_exec": 0.06, "u_spec": 0.05, "u_risk": 0.08}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 2 |
+
{"mode": "low_uncertainty", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: multiple corroborating sources agree and carry stable provenance.\nValidation state: latest validator receipt succeeded without rollback.\nWorktree state: clean snapshot with no conflicting edit.\nPatch state: none pending.\nToolspec state: bounded call contract is fully aligned.", "target": {"u_answer": 0.08, "u_evidence": 0.1, "u_exec": 0.06, "u_spec": 0.05, "u_risk": 0.08}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 3 |
+
{"mode": "low_uncertainty", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: multiple corroborating sources agree and carry stable provenance.\nValidation state: latest validator receipt succeeded without rollback.\nWorktree state: clean snapshot with no conflicting edit.\nPatch state: none pending.\nToolspec state: bounded call contract is fully aligned.", "target": {"u_answer": 0.08, "u_evidence": 0.1, "u_exec": 0.06, "u_spec": 0.05, "u_risk": 0.08}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 4 |
+
{"mode": "low_uncertainty", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: decide whether a post-generation self-check loop is still required.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: multiple corroborating sources agree and carry stable provenance.\nValidation state: latest validator receipt succeeded without rollback.\nWorktree state: clean snapshot with no conflicting edit.\nPatch state: none pending.\nToolspec state: bounded call contract is fully aligned.", "target": {"u_answer": 0.08, "u_evidence": 0.1, "u_exec": 0.06, "u_spec": 0.05, "u_risk": 0.08}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 5 |
+
{"mode": "low_uncertainty", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: multiple corroborating sources agree and carry stable provenance.\nValidation state: latest validator receipt succeeded without rollback.\nWorktree state: clean snapshot with no conflicting edit.\nPatch state: none pending.\nToolspec state: bounded call contract is fully aligned.", "target": {"u_answer": 0.08, "u_evidence": 0.1, "u_exec": 0.06, "u_spec": 0.05, "u_risk": 0.08}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
| 6 |
+
{"mode": "evidence_gap", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: retrieval produced sparse evidence and citation confidence remains low.\nValidation state: no explicit failure yet, but claim support is visibly underdetermined.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: currently legal.", "target": {"u_answer": 0.42, "u_evidence": 0.86, "u_exec": 0.38, "u_spec": 0.18, "u_risk": 0.58}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 7 |
+
{"mode": "evidence_gap", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: retrieval produced sparse evidence and citation confidence remains low.\nValidation state: no explicit failure yet, but claim support is visibly underdetermined.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: currently legal.", "target": {"u_answer": 0.42, "u_evidence": 0.86, "u_exec": 0.38, "u_spec": 0.18, "u_risk": 0.58}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 8 |
+
{"mode": "evidence_gap", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: retrieval produced sparse evidence and citation confidence remains low.\nValidation state: no explicit failure yet, but claim support is visibly underdetermined.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: currently legal.", "target": {"u_answer": 0.42, "u_evidence": 0.86, "u_exec": 0.38, "u_spec": 0.18, "u_risk": 0.58}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 9 |
+
{"mode": "evidence_gap", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: retrieval produced sparse evidence and citation confidence remains low.\nValidation state: no explicit failure yet, but claim support is visibly underdetermined.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: currently legal.", "target": {"u_answer": 0.42, "u_evidence": 0.86, "u_exec": 0.38, "u_spec": 0.18, "u_risk": 0.58}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 10 |
+
{"mode": "evidence_gap", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: estimate runtime risk before trusting a partially-validated repository change.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: retrieval produced sparse evidence and citation confidence remains low.\nValidation state: no explicit failure yet, but claim support is visibly underdetermined.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: currently legal.", "target": {"u_answer": 0.42, "u_evidence": 0.86, "u_exec": 0.38, "u_spec": 0.18, "u_risk": 0.58}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
| 11 |
+
{"mode": "exec_required", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: reasoning narrows the options but cannot finalize action legality.\nValidation state: no execution receipt exists yet.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: valid once execution is routed.", "target": {"u_answer": 0.48, "u_evidence": 0.52, "u_exec": 0.94, "u_spec": 0.16, "u_risk": 0.55}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 12 |
+
{"mode": "exec_required", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: decide whether a post-generation self-check loop is still required.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: reasoning narrows the options but cannot finalize action legality.\nValidation state: no execution receipt exists yet.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: valid once execution is routed.", "target": {"u_answer": 0.48, "u_evidence": 0.52, "u_exec": 0.94, "u_spec": 0.16, "u_risk": 0.55}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 13 |
+
{"mode": "exec_required", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: estimate runtime risk before trusting a partially-validated repository change.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: reasoning narrows the options but cannot finalize action legality.\nValidation state: no execution receipt exists yet.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: valid once execution is routed.", "target": {"u_answer": 0.48, "u_evidence": 0.52, "u_exec": 0.94, "u_spec": 0.16, "u_risk": 0.55}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 14 |
+
{"mode": "exec_required", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: decide whether a post-generation self-check loop is still required.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: reasoning narrows the options but cannot finalize action legality.\nValidation state: no execution receipt exists yet.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: valid once execution is routed.", "target": {"u_answer": 0.48, "u_evidence": 0.52, "u_exec": 0.94, "u_spec": 0.16, "u_risk": 0.55}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 15 |
+
{"mode": "exec_required", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: reasoning narrows the options but cannot finalize action legality.\nValidation state: no execution receipt exists yet.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: valid once execution is routed.", "target": {"u_answer": 0.48, "u_evidence": 0.52, "u_exec": 0.94, "u_spec": 0.16, "u_risk": 0.55}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
| 16 |
+
{"mode": "spec_mismatch", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: decide whether a post-generation self-check loop is still required.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: answer content is mostly grounded.\nValidation state: validator rejects the call because required fields do not match the live toolspec.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: contract mismatch remains unresolved.", "target": {"u_answer": 0.34, "u_evidence": 0.28, "u_exec": 0.74, "u_spec": 0.96, "u_risk": 0.62}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 17 |
+
{"mode": "spec_mismatch", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: estimate runtime risk before trusting a partially-validated repository change.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: answer content is mostly grounded.\nValidation state: validator rejects the call because required fields do not match the live toolspec.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: contract mismatch remains unresolved.", "target": {"u_answer": 0.34, "u_evidence": 0.28, "u_exec": 0.74, "u_spec": 0.96, "u_risk": 0.62}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 18 |
+
{"mode": "spec_mismatch", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: answer content is mostly grounded.\nValidation state: validator rejects the call because required fields do not match the live toolspec.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: contract mismatch remains unresolved.", "target": {"u_answer": 0.34, "u_evidence": 0.28, "u_exec": 0.74, "u_spec": 0.96, "u_risk": 0.62}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 19 |
+
{"mode": "spec_mismatch", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: answer content is mostly grounded.\nValidation state: validator rejects the call because required fields do not match the live toolspec.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: contract mismatch remains unresolved.", "target": {"u_answer": 0.34, "u_evidence": 0.28, "u_exec": 0.74, "u_spec": 0.96, "u_risk": 0.62}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 20 |
+
{"mode": "spec_mismatch", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: answer content is mostly grounded.\nValidation state: validator rejects the call because required fields do not match the live toolspec.\nWorktree state: unchanged.\nPatch state: none.\nToolspec state: contract mismatch remains unresolved.", "target": {"u_answer": 0.34, "u_evidence": 0.28, "u_exec": 0.74, "u_spec": 0.96, "u_risk": 0.62}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
| 21 |
+
{"mode": "high_risk", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: only moderate support is present.\nValidation state: no execution or validator adjudication has reduced the risk yet.\nWorktree state: unchanged.\nPatch state: none.\nRisk note: unsupported output would cause material downstream harm.", "target": {"u_answer": 0.6, "u_evidence": 0.7, "u_exec": 0.48, "u_spec": 0.3, "u_risk": 0.99}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 22 |
+
{"mode": "high_risk", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: only moderate support is present.\nValidation state: no execution or validator adjudication has reduced the risk yet.\nWorktree state: unchanged.\nPatch state: none.\nRisk note: unsupported output would cause material downstream harm.", "target": {"u_answer": 0.6, "u_evidence": 0.7, "u_exec": 0.48, "u_spec": 0.3, "u_risk": 0.99}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 23 |
+
{"mode": "high_risk", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: decide whether a post-generation self-check loop is still required.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: only moderate support is present.\nValidation state: no execution or validator adjudication has reduced the risk yet.\nWorktree state: unchanged.\nPatch state: none.\nRisk note: unsupported output would cause material downstream harm.", "target": {"u_answer": 0.6, "u_evidence": 0.7, "u_exec": 0.48, "u_spec": 0.3, "u_risk": 0.99}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 24 |
+
{"mode": "high_risk", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: only moderate support is present.\nValidation state: no execution or validator adjudication has reduced the risk yet.\nWorktree state: unchanged.\nPatch state: none.\nRisk note: unsupported output would cause material downstream harm.", "target": {"u_answer": 0.6, "u_evidence": 0.7, "u_exec": 0.48, "u_spec": 0.3, "u_risk": 0.99}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 25 |
+
{"mode": "high_risk", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: only moderate support is present.\nValidation state: no execution or validator adjudication has reduced the risk yet.\nWorktree state: unchanged.\nPatch state: none.\nRisk note: unsupported output would cause material downstream harm.", "target": {"u_answer": 0.6, "u_evidence": 0.7, "u_exec": 0.48, "u_spec": 0.3, "u_risk": 0.99}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
| 26 |
+
{"mode": "conflicting_evidence", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: one source supports the claim while another equally credible source negates it.\nValidation state: validator has not broken the tie.\nWorktree state: unchanged.\nPatch state: none.\nConflict note: evidence graph still contains unresolved contradiction edges.", "target": {"u_answer": 0.78, "u_evidence": 0.82, "u_exec": 0.46, "u_spec": 0.22, "u_risk": 0.74}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 27 |
+
{"mode": "conflicting_evidence", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: one source supports the claim while another equally credible source negates it.\nValidation state: validator has not broken the tie.\nWorktree state: unchanged.\nPatch state: none.\nConflict note: evidence graph still contains unresolved contradiction edges.", "target": {"u_answer": 0.78, "u_evidence": 0.82, "u_exec": 0.46, "u_spec": 0.22, "u_risk": 0.74}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 28 |
+
{"mode": "conflicting_evidence", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: estimate runtime risk before trusting a partially-validated repository change.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: one source supports the claim while another equally credible source negates it.\nValidation state: validator has not broken the tie.\nWorktree state: unchanged.\nPatch state: none.\nConflict note: evidence graph still contains unresolved contradiction edges.", "target": {"u_answer": 0.78, "u_evidence": 0.82, "u_exec": 0.46, "u_spec": 0.22, "u_risk": 0.74}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 29 |
+
{"mode": "conflicting_evidence", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: one source supports the claim while another equally credible source negates it.\nValidation state: validator has not broken the tie.\nWorktree state: unchanged.\nPatch state: none.\nConflict note: evidence graph still contains unresolved contradiction edges.", "target": {"u_answer": 0.78, "u_evidence": 0.82, "u_exec": 0.46, "u_spec": 0.22, "u_risk": 0.74}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 30 |
+
{"mode": "conflicting_evidence", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: one source supports the claim while another equally credible source negates it.\nValidation state: validator has not broken the tie.\nWorktree state: unchanged.\nPatch state: none.\nConflict note: evidence graph still contains unresolved contradiction edges.", "target": {"u_answer": 0.78, "u_evidence": 0.82, "u_exec": 0.46, "u_spec": 0.22, "u_risk": 0.74}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
| 31 |
+
{"mode": "validator_negation", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: previous support is now contested by a failing validator receipt.\nValidation state: latest validator explicitly rejects the active state.\nWorktree state: patch remains mounted.\nPatch state: rollback or repair decision is still pending.\nFailure route: rollback bridge is armed.", "target": {"u_answer": 0.86, "u_evidence": 0.64, "u_exec": 0.72, "u_spec": 0.2, "u_risk": 0.8}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 32 |
+
{"mode": "validator_negation", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: previous support is now contested by a failing validator receipt.\nValidation state: latest validator explicitly rejects the active state.\nWorktree state: patch remains mounted.\nPatch state: rollback or repair decision is still pending.\nFailure route: rollback bridge is armed.", "target": {"u_answer": 0.86, "u_evidence": 0.64, "u_exec": 0.72, "u_spec": 0.2, "u_risk": 0.8}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 33 |
+
{"mode": "validator_negation", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: estimate runtime risk before trusting a partially-validated repository change.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: previous support is now contested by a failing validator receipt.\nValidation state: latest validator explicitly rejects the active state.\nWorktree state: patch remains mounted.\nPatch state: rollback or repair decision is still pending.\nFailure route: rollback bridge is armed.", "target": {"u_answer": 0.86, "u_evidence": 0.64, "u_exec": 0.72, "u_spec": 0.2, "u_risk": 0.8}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 34 |
+
{"mode": "validator_negation", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: decide whether a post-generation self-check loop is still required.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: previous support is now contested by a failing validator receipt.\nValidation state: latest validator explicitly rejects the active state.\nWorktree state: patch remains mounted.\nPatch state: rollback or repair decision is still pending.\nFailure route: rollback bridge is armed.", "target": {"u_answer": 0.86, "u_evidence": 0.64, "u_exec": 0.72, "u_spec": 0.2, "u_risk": 0.8}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 35 |
+
{"mode": "validator_negation", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: decide whether a post-generation self-check loop is still required.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: previous support is now contested by a failing validator receipt.\nValidation state: latest validator explicitly rejects the active state.\nWorktree state: patch remains mounted.\nPatch state: rollback or repair decision is still pending.\nFailure route: rollback bridge is armed.", "target": {"u_answer": 0.86, "u_evidence": 0.64, "u_exec": 0.72, "u_spec": 0.2, "u_risk": 0.8}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
| 36 |
+
{"mode": "worktree_conflict", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: moderate support.\nValidation state: validator cannot finalize because the workspace is not in a coherent state.\nWorktree state: overlapping edits violate expected snapshot continuity.\nPatch state: pending.\nPermission note: bounded write discipline is still active.", "target": {"u_answer": 0.44, "u_evidence": 0.34, "u_exec": 0.7, "u_spec": 0.78, "u_risk": 0.68}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 37 |
+
{"mode": "worktree_conflict", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: moderate support.\nValidation state: validator cannot finalize because the workspace is not in a coherent state.\nWorktree state: overlapping edits violate expected snapshot continuity.\nPatch state: pending.\nPermission note: bounded write discipline is still active.", "target": {"u_answer": 0.44, "u_evidence": 0.34, "u_exec": 0.7, "u_spec": 0.78, "u_risk": 0.68}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 38 |
+
{"mode": "worktree_conflict", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: moderate support.\nValidation state: validator cannot finalize because the workspace is not in a coherent state.\nWorktree state: overlapping edits violate expected snapshot continuity.\nPatch state: pending.\nPermission note: bounded write discipline is still active.", "target": {"u_answer": 0.44, "u_evidence": 0.34, "u_exec": 0.7, "u_spec": 0.78, "u_risk": 0.68}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 39 |
+
{"mode": "worktree_conflict", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: moderate support.\nValidation state: validator cannot finalize because the workspace is not in a coherent state.\nWorktree state: overlapping edits violate expected snapshot continuity.\nPatch state: pending.\nPermission note: bounded write discipline is still active.", "target": {"u_answer": 0.44, "u_evidence": 0.34, "u_exec": 0.7, "u_spec": 0.78, "u_risk": 0.68}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 40 |
+
{"mode": "worktree_conflict", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: estimate runtime risk before trusting a partially-validated repository change.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: moderate support.\nValidation state: validator cannot finalize because the workspace is not in a coherent state.\nWorktree state: overlapping edits violate expected snapshot continuity.\nPatch state: pending.\nPermission note: bounded write discipline is still active.", "target": {"u_answer": 0.44, "u_evidence": 0.34, "u_exec": 0.7, "u_spec": 0.78, "u_risk": 0.68}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
| 41 |
+
{"mode": "patch_pending", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: estimate runtime risk before trusting a partially-validated repository change.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: adequate for a repair attempt but not yet for final acceptance.\nValidation state: validator requested one more bounded patch cycle.\nWorktree state: stable.\nPatch state: corrective artifact exists but remains unverified.\nReceipt state: final acceptance receipt not yet produced.", "target": {"u_answer": 0.38, "u_evidence": 0.36, "u_exec": 0.66, "u_spec": 0.26, "u_risk": 0.5}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 42 |
+
{"mode": "patch_pending", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: adequate for a repair attempt but not yet for final acceptance.\nValidation state: validator requested one more bounded patch cycle.\nWorktree state: stable.\nPatch state: corrective artifact exists but remains unverified.\nReceipt state: final acceptance receipt not yet produced.", "target": {"u_answer": 0.38, "u_evidence": 0.36, "u_exec": 0.66, "u_spec": 0.26, "u_risk": 0.5}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 43 |
+
{"mode": "patch_pending", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: adequate for a repair attempt but not yet for final acceptance.\nValidation state: validator requested one more bounded patch cycle.\nWorktree state: stable.\nPatch state: corrective artifact exists but remains unverified.\nReceipt state: final acceptance receipt not yet produced.", "target": {"u_answer": 0.38, "u_evidence": 0.36, "u_exec": 0.66, "u_spec": 0.26, "u_risk": 0.5}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 44 |
+
{"mode": "patch_pending", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: adequate for a repair attempt but not yet for final acceptance.\nValidation state: validator requested one more bounded patch cycle.\nWorktree state: stable.\nPatch state: corrective artifact exists but remains unverified.\nReceipt state: final acceptance receipt not yet produced.", "target": {"u_answer": 0.38, "u_evidence": 0.36, "u_exec": 0.66, "u_spec": 0.26, "u_risk": 0.5}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 45 |
+
{"mode": "patch_pending", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: adequate for a repair attempt but not yet for final acceptance.\nValidation state: validator requested one more bounded patch cycle.\nWorktree state: stable.\nPatch state: corrective artifact exists but remains unverified.\nReceipt state: final acceptance receipt not yet produced.", "target": {"u_answer": 0.38, "u_evidence": 0.36, "u_exec": 0.66, "u_spec": 0.26, "u_risk": 0.5}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
| 46 |
+
{"mode": "self_check_failure", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: estimate runtime risk before trusting a partially-validated repository change.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: prior reasoning looked sound.\nValidation state: the self-check loop surfaced a failing test after generation.\nWorktree state: stable.\nPatch state: last edit remains mounted.\nInterpretation: submit would be premature without another bounded repair loop.", "target": {"u_answer": 0.55, "u_evidence": 0.3, "u_exec": 0.88, "u_spec": 0.22, "u_risk": 0.63}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 47 |
+
{"mode": "self_check_failure", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: prior reasoning looked sound.\nValidation state: the self-check loop surfaced a failing test after generation.\nWorktree state: stable.\nPatch state: last edit remains mounted.\nInterpretation: submit would be premature without another bounded repair loop.", "target": {"u_answer": 0.55, "u_evidence": 0.3, "u_exec": 0.88, "u_spec": 0.22, "u_risk": 0.63}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 48 |
+
{"mode": "self_check_failure", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: decide whether a post-generation self-check loop is still required.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: prior reasoning looked sound.\nValidation state: the self-check loop surfaced a failing test after generation.\nWorktree state: stable.\nPatch state: last edit remains mounted.\nInterpretation: submit would be premature without another bounded repair loop.", "target": {"u_answer": 0.55, "u_evidence": 0.3, "u_exec": 0.88, "u_spec": 0.22, "u_risk": 0.63}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 49 |
+
{"mode": "self_check_failure", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: estimate runtime risk before trusting a partially-validated repository change.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: prior reasoning looked sound.\nValidation state: the self-check loop surfaced a failing test after generation.\nWorktree state: stable.\nPatch state: last edit remains mounted.\nInterpretation: submit would be premature without another bounded repair loop.", "target": {"u_answer": 0.55, "u_evidence": 0.3, "u_exec": 0.88, "u_spec": 0.22, "u_risk": 0.63}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 50 |
+
{"mode": "self_check_failure", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: prior reasoning looked sound.\nValidation state: the self-check loop surfaced a failing test after generation.\nWorktree state: stable.\nPatch state: last edit remains mounted.\nInterpretation: submit would be premature without another bounded repair loop.", "target": {"u_answer": 0.55, "u_evidence": 0.3, "u_exec": 0.88, "u_spec": 0.22, "u_risk": 0.63}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
| 51 |
+
{"mode": "reverse_engineering_ambiguity", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #1.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: repository topology is only partially reconstructed.\nValidation state: no sandbox failure exists yet, but architecture intent is not resolved.\nWorktree state: unchanged.\nPatch state: none.\nInterpretation: reverse-engineering readiness is required before legal mutation.", "target": {"u_answer": 0.52, "u_evidence": 0.76, "u_exec": 0.68, "u_spec": 0.72, "u_risk": 0.44}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 0}}
|
| 52 |
+
{"mode": "reverse_engineering_ambiguity", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #2.\nTask brief: decide whether another sandbox run is needed before accepting a patch.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: repository topology is only partially reconstructed.\nValidation state: no sandbox failure exists yet, but architecture intent is not resolved.\nWorktree state: unchanged.\nPatch state: none.\nInterpretation: reverse-engineering readiness is required before legal mutation.", "target": {"u_answer": 0.52, "u_evidence": 0.76, "u_exec": 0.68, "u_spec": 0.72, "u_risk": 0.44}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 1}}
|
| 53 |
+
{"mode": "reverse_engineering_ambiguity", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #3.\nTask brief: infer repository intent when the codebase must be reverse engineered before editing.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: repository topology is only partially reconstructed.\nValidation state: no sandbox failure exists yet, but architecture intent is not resolved.\nWorktree state: unchanged.\nPatch state: none.\nInterpretation: reverse-engineering readiness is required before legal mutation.", "target": {"u_answer": 0.52, "u_evidence": 0.76, "u_exec": 0.68, "u_spec": 0.72, "u_risk": 0.44}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 2}}
|
| 54 |
+
{"mode": "reverse_engineering_ambiguity", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #4.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: repository topology is only partially reconstructed.\nValidation state: no sandbox failure exists yet, but architecture intent is not resolved.\nWorktree state: unchanged.\nPatch state: none.\nInterpretation: reverse-engineering readiness is required before legal mutation.", "target": {"u_answer": 0.52, "u_evidence": 0.76, "u_exec": 0.68, "u_spec": 0.72, "u_risk": 0.44}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 3}}
|
| 55 |
+
{"mode": "reverse_engineering_ambiguity", "product_line": "veriloop_coder", "prompt": "Evaluation scenario #5.\nTask brief: judge whether a code patch is ready to submit after validator feedback.\nProduct line: veriloop_coder.\nRuntime mode: verify-before-submit.\nHarness state: validator loop active; permission context active; worktree state tracked; runtime protocol active.\nReturn uncertainty only as a structured five-dimensional estimate.\nEvidence bundle: repository topology is only partially reconstructed.\nValidation state: no sandbox failure exists yet, but architecture intent is not resolved.\nWorktree state: unchanged.\nPatch state: none.\nInterpretation: reverse-engineering readiness is required before legal mutation.", "target": {"u_answer": 0.52, "u_evidence": 0.76, "u_exec": 0.68, "u_spec": 0.72, "u_risk": 0.44}, "metadata": {"synthetic": true, "split": "eval", "sample_idx": 4}}
|
uncertainty_adapter/uncertainty_head.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bdd4782fe5d2f26c184c8690a07527cffa99eebe763e415a4ef449ed8046c88a
|
| 3 |
+
size 25220609
|
uncertainty_adapter/uncertainty_train.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
uncertainty_adapter/uncertainty_training_manifest.json
ADDED
|
@@ -0,0 +1,105 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"adapter_exported": true,
|
| 3 |
+
"auto_lora_from_ia3": false,
|
| 4 |
+
"best_epoch": 2,
|
| 5 |
+
"best_quality_score": 0.8625935807221907,
|
| 6 |
+
"dataset_summary": {
|
| 7 |
+
"eval_modes": {
|
| 8 |
+
"conflicting_evidence": 5,
|
| 9 |
+
"evidence_gap": 5,
|
| 10 |
+
"exec_required": 5,
|
| 11 |
+
"high_risk": 5,
|
| 12 |
+
"low_uncertainty": 5,
|
| 13 |
+
"patch_pending": 5,
|
| 14 |
+
"reverse_engineering_ambiguity": 5,
|
| 15 |
+
"self_check_failure": 5,
|
| 16 |
+
"spec_mismatch": 5,
|
| 17 |
+
"validator_negation": 5,
|
| 18 |
+
"worktree_conflict": 5
|
| 19 |
+
},
|
| 20 |
+
"eval_size": 55,
|
| 21 |
+
"train_modes": {
|
| 22 |
+
"conflicting_evidence": 14,
|
| 23 |
+
"evidence_gap": 14,
|
| 24 |
+
"exec_required": 14,
|
| 25 |
+
"high_risk": 14,
|
| 26 |
+
"low_uncertainty": 14,
|
| 27 |
+
"patch_pending": 14,
|
| 28 |
+
"reverse_engineering_ambiguity": 14,
|
| 29 |
+
"self_check_failure": 14,
|
| 30 |
+
"spec_mismatch": 14,
|
| 31 |
+
"validator_negation": 14,
|
| 32 |
+
"worktree_conflict": 14
|
| 33 |
+
},
|
| 34 |
+
"train_size": 154
|
| 35 |
+
},
|
| 36 |
+
"epochs_completed": 4,
|
| 37 |
+
"eval_metrics": {
|
| 38 |
+
"adapter_exported": true,
|
| 39 |
+
"auto_lora_from_ia3": false,
|
| 40 |
+
"best_epoch": 2,
|
| 41 |
+
"best_quality_score": 0.8625935807221907,
|
| 42 |
+
"count": 55,
|
| 43 |
+
"mae": {
|
| 44 |
+
"u_answer": 0.15174226462841034,
|
| 45 |
+
"u_evidence": 0.19610758125782013,
|
| 46 |
+
"u_exec": 0.18561214208602905,
|
| 47 |
+
"u_risk": 0.1553734689950943,
|
| 48 |
+
"u_spec": 0.21633382141590118
|
| 49 |
+
},
|
| 50 |
+
"mean_mae": 0.18103384971618652,
|
| 51 |
+
"mean_rmse": 0.24169571697711945,
|
| 52 |
+
"moderate_accuracy": 0.6727272727272727,
|
| 53 |
+
"peft_method": "lora_narrow",
|
| 54 |
+
"quality_score": 0.8625935807221907,
|
| 55 |
+
"rmse": {
|
| 56 |
+
"u_answer": 0.18724055588245392,
|
| 57 |
+
"u_evidence": 0.22527915239334106,
|
| 58 |
+
"u_exec": 0.25238174200057983,
|
| 59 |
+
"u_risk": 0.20667441189289093,
|
| 60 |
+
"u_spec": 0.3369026482105255
|
| 61 |
+
},
|
| 62 |
+
"tight_accuracy": 0.4,
|
| 63 |
+
"used_peft": true,
|
| 64 |
+
"weighted_mae": 0.18083095811830807,
|
| 65 |
+
"weighted_rmse": 0.24125460771003793
|
| 66 |
+
},
|
| 67 |
+
"load_meta": {
|
| 68 |
+
"chosen_class": "AutoModelForCausalLM",
|
| 69 |
+
"hidden_size": 2048,
|
| 70 |
+
"quantization_mode": "4bit"
|
| 71 |
+
},
|
| 72 |
+
"peft_method": "lora_narrow",
|
| 73 |
+
"requested_method": "lora_narrow",
|
| 74 |
+
"requested_target_modules": [
|
| 75 |
+
"uncertainty_head",
|
| 76 |
+
"uncertainty_head.calibration_mlp",
|
| 77 |
+
"uncertainty_head.proj"
|
| 78 |
+
],
|
| 79 |
+
"resolved_target_modules": [
|
| 80 |
+
"surface_host.uncertainty_head.calibration_mlp",
|
| 81 |
+
"surface_host.uncertainty_head.proj"
|
| 82 |
+
],
|
| 83 |
+
"selected_target_modules": [
|
| 84 |
+
"uncertainty_head",
|
| 85 |
+
"uncertainty_head.calibration_mlp",
|
| 86 |
+
"uncertainty_head.proj"
|
| 87 |
+
],
|
| 88 |
+
"status": "trained",
|
| 89 |
+
"train_metrics": {
|
| 90 |
+
"adapter_exported": true,
|
| 91 |
+
"auto_lora_from_ia3": false,
|
| 92 |
+
"best_epoch": 2,
|
| 93 |
+
"best_quality_score": 0.8625935807221907,
|
| 94 |
+
"epochs_completed": 4,
|
| 95 |
+
"loss": 0.009006613283418119,
|
| 96 |
+
"micro_batches": 154,
|
| 97 |
+
"micro_batches_total": 616,
|
| 98 |
+
"optimizer_steps": 10,
|
| 99 |
+
"optimizer_steps_total": 40,
|
| 100 |
+
"peft_method": "lora_narrow",
|
| 101 |
+
"used_peft": true
|
| 102 |
+
},
|
| 103 |
+
"unresolved_target_modules": [],
|
| 104 |
+
"used_peft": true
|
| 105 |
+
}
|