Update README.md
Browse files
README.md
CHANGED
|
@@ -1,63 +1,183 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
-
|
| 5 |
-
-
|
| 6 |
-
-
|
| 7 |
-
-
|
| 8 |
-
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
-
|
| 15 |
-
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
|
| 60 |
-
|
|
| 61 |
-
|
|
| 62 |
-
|
|
| 63 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- merge
|
| 4 |
+
- mergekit
|
| 5 |
+
- lazymergekit
|
| 6 |
+
- flemmingmiguel/NeuDist-Ro-7B
|
| 7 |
+
- johannhartmann/Brezn3
|
| 8 |
+
- ResplendentAI/Flora_DPO_7B
|
| 9 |
+
base_model:
|
| 10 |
+
- flemmingmiguel/NeuDist-Ro-7B
|
| 11 |
+
- johannhartmann/Brezn3
|
| 12 |
+
- ResplendentAI/Flora_DPO_7B
|
| 13 |
+
language:
|
| 14 |
+
- de
|
| 15 |
+
- en
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# Spaetzle-v8-7b
|
| 19 |
+
|
| 20 |
+
This model is supposed to show adequate performance in German and English on a number of tasks, while mostly behaving well, that is, without rambling on, intermixing tokens from different templates in training and adapting, etc.
|
| 21 |
+
|
| 22 |
+
It is mostly a quick test, and considerably weaker in German grammar and orthography than DiscoLM e.g., but for use cases where this is not too important, but e.g. instruction following, reasoning, etc, it might actually be a little bit preferable.
|
| 23 |
+
|
| 24 |
+
It is a merge of the following models using [LazyMergekit](https://colab.research.google.com/drive/1obulZ1ROXHjYLn6PPZJwRR6GzgQogxxb?usp=sharing):
|
| 25 |
+
* [flemmingmiguel/NeuDist-Ro-7B](https://huggingface.co/flemmingmiguel/NeuDist-Ro-7B)
|
| 26 |
+
* [johannhartmann/Brezn3](https://huggingface.co/johannhartmann/Brezn3)
|
| 27 |
+
* [ResplendentAI/Flora_DPO_7B](https://huggingface.co/ResplendentAI/Flora_DPO_7B)
|
| 28 |
+
* on the basis of [mayflowergmbh/Wiedervereinigung-7b-dpo-laser](https://huggingface.co/mayflowergmbh/Wiedervereinigung-7b-dpo-laser)
|
| 29 |
+
|
| 30 |
+
All credits are due to the creators of those original models and the training datasets involved.
|
| 31 |
+
|
| 32 |
+
For a suitable quantized version, try [cstr/Spaetzle-v8-7b-GGUF](https://huggingface.co/cstr/Spaetzle-v8-7b-GGUF)
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
## Evaluation
|
| 36 |
+
[Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
|
| 37 |
+
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_cstr__Spaetzle-v8-7b)
|
| 38 |
+
|
| 39 |
+
| Metric |Value|
|
| 40 |
+
|---------------------------------|----:|
|
| 41 |
+
|Avg. |72.27|
|
| 42 |
+
|AI2 Reasoning Challenge (25-Shot)|68.69|
|
| 43 |
+
|HellaSwag (10-Shot) |86.68|
|
| 44 |
+
|MMLU (5-Shot) |64.60|
|
| 45 |
+
|TruthfulQA (0-shot) |64.05|
|
| 46 |
+
|Winogrande (5-shot) |81.45|
|
| 47 |
+
|GSM8k (5-shot) |68.16|
|
| 48 |
+
|
| 49 |
+
EQ-Bench (v2_de): 61.04 / english (v2): 78.3
|
| 50 |
+
|
| 51 |
+
| Model |AGIEval|GPT4All|TruthfulQA|Bigbench|Average|
|
| 52 |
+
|------------------------------------------------------------|------:|------:|---------:|-------:|------:|
|
| 53 |
+
|[Spaetzle-v8-7b](https://huggingface.co/cstr/Spaetzle-v8-7b)| 45.31| 75.69| 63.94| 45.57| 57.63|
|
| 54 |
+
|
| 55 |
+
### AGIEval
|
| 56 |
+
| Task |Version| Metric |Value| |Stderr|
|
| 57 |
+
|------------------------------|------:|--------|----:|---|-----:|
|
| 58 |
+
|agieval_aqua_rat | 0|acc |25.59|± | 2.74|
|
| 59 |
+
| | |acc_norm|24.80|± | 2.72|
|
| 60 |
+
|agieval_logiqa_en | 0|acc |39.63|± | 1.92|
|
| 61 |
+
| | |acc_norm|39.78|± | 1.92|
|
| 62 |
+
|agieval_lsat_ar | 0|acc |23.48|± | 2.80|
|
| 63 |
+
| | |acc_norm|24.35|± | 2.84|
|
| 64 |
+
|agieval_lsat_lr | 0|acc |50.98|± | 2.22|
|
| 65 |
+
| | |acc_norm|51.96|± | 2.21|
|
| 66 |
+
|agieval_lsat_rc | 0|acc |62.08|± | 2.96|
|
| 67 |
+
| | |acc_norm|62.83|± | 2.95|
|
| 68 |
+
|agieval_sat_en | 0|acc |78.64|± | 2.86|
|
| 69 |
+
| | |acc_norm|79.13|± | 2.84|
|
| 70 |
+
|agieval_sat_en_without_passage| 0|acc |44.66|± | 3.47|
|
| 71 |
+
| | |acc_norm|44.66|± | 3.47|
|
| 72 |
+
|agieval_sat_math | 0|acc |37.27|± | 3.27|
|
| 73 |
+
| | |acc_norm|35.00|± | 3.22|
|
| 74 |
+
|
| 75 |
+
Average: 45.31%
|
| 76 |
+
|
| 77 |
+
### GPT4All
|
| 78 |
+
| Task |Version| Metric |Value| |Stderr|
|
| 79 |
+
|-------------|------:|--------|----:|---|-----:|
|
| 80 |
+
|arc_challenge| 0|acc |63.14|± | 1.41|
|
| 81 |
+
| | |acc_norm|64.51|± | 1.40|
|
| 82 |
+
|arc_easy | 0|acc |85.98|± | 0.71|
|
| 83 |
+
| | |acc_norm|82.49|± | 0.78|
|
| 84 |
+
|boolq | 1|acc |88.10|± | 0.57|
|
| 85 |
+
|hellaswag | 0|acc |66.31|± | 0.47|
|
| 86 |
+
| | |acc_norm|85.17|± | 0.35|
|
| 87 |
+
|openbookqa | 0|acc |38.00|± | 2.17|
|
| 88 |
+
| | |acc_norm|47.20|± | 2.23|
|
| 89 |
+
|piqa | 0|acc |83.35|± | 0.87|
|
| 90 |
+
| | |acc_norm|84.17|± | 0.85|
|
| 91 |
+
|winogrande | 0|acc |78.22|± | 1.16|
|
| 92 |
+
|
| 93 |
+
Average: 75.69%
|
| 94 |
+
|
| 95 |
+
### TruthfulQA
|
| 96 |
+
| Task |Version|Metric|Value| |Stderr|
|
| 97 |
+
|-------------|------:|------|----:|---|-----:|
|
| 98 |
+
|truthfulqa_mc| 1|mc1 |47.74|± | 1.75|
|
| 99 |
+
| | |mc2 |63.94|± | 1.53|
|
| 100 |
+
|
| 101 |
+
Average: 63.94%
|
| 102 |
+
|
| 103 |
+
### Bigbench
|
| 104 |
+
| Task |Version| Metric |Value| |Stderr|
|
| 105 |
+
|------------------------------------------------|------:|---------------------|----:|---|-----:|
|
| 106 |
+
|bigbench_causal_judgement | 0|multiple_choice_grade|56.84|± | 3.60|
|
| 107 |
+
|bigbench_date_understanding | 0|multiple_choice_grade|66.12|± | 2.47|
|
| 108 |
+
|bigbench_disambiguation_qa | 0|multiple_choice_grade|41.47|± | 3.07|
|
| 109 |
+
|bigbench_geometric_shapes | 0|multiple_choice_grade|22.01|± | 2.19|
|
| 110 |
+
| | |exact_str_match | 0.00|± | 0.00|
|
| 111 |
+
|bigbench_logical_deduction_five_objects | 0|multiple_choice_grade|31.40|± | 2.08|
|
| 112 |
+
|bigbench_logical_deduction_seven_objects | 0|multiple_choice_grade|23.14|± | 1.60|
|
| 113 |
+
|bigbench_logical_deduction_three_objects | 0|multiple_choice_grade|56.00|± | 2.87|
|
| 114 |
+
|bigbench_movie_recommendation | 0|multiple_choice_grade|45.00|± | 2.23|
|
| 115 |
+
|bigbench_navigate | 0|multiple_choice_grade|50.70|± | 1.58|
|
| 116 |
+
|bigbench_reasoning_about_colored_objects | 0|multiple_choice_grade|70.05|± | 1.02|
|
| 117 |
+
|bigbench_ruin_names | 0|multiple_choice_grade|45.54|± | 2.36|
|
| 118 |
+
|bigbench_salient_translation_error_detection | 0|multiple_choice_grade|26.05|± | 1.39|
|
| 119 |
+
|bigbench_snarks | 0|multiple_choice_grade|71.82|± | 3.35|
|
| 120 |
+
|bigbench_sports_understanding | 0|multiple_choice_grade|72.92|± | 1.42|
|
| 121 |
+
|bigbench_temporal_sequences | 0|multiple_choice_grade|44.20|± | 1.57|
|
| 122 |
+
|bigbench_tracking_shuffled_objects_five_objects | 0|multiple_choice_grade|22.80|± | 1.19|
|
| 123 |
+
|bigbench_tracking_shuffled_objects_seven_objects| 0|multiple_choice_grade|18.23|± | 0.92|
|
| 124 |
+
|bigbench_tracking_shuffled_objects_three_objects| 0|multiple_choice_grade|56.00|± | 2.87|
|
| 125 |
+
|
| 126 |
+
Average: 45.57%
|
| 127 |
+
|
| 128 |
+
Average score: 57.63%
|
| 129 |
+
|
| 130 |
+
## 💻 Usage
|
| 131 |
+
|
| 132 |
+
```python
|
| 133 |
+
!pip install -qU transformers accelerate
|
| 134 |
+
|
| 135 |
+
from transformers import AutoTokenizer
|
| 136 |
+
import transformers
|
| 137 |
+
import torch
|
| 138 |
+
|
| 139 |
+
model = "cstr/Spaetzle-v8-7b"
|
| 140 |
+
messages = [{"role": "user", "content": "What is a large language model?"}]
|
| 141 |
+
|
| 142 |
+
tokenizer = AutoTokenizer.from_pretrained(model)
|
| 143 |
+
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 144 |
+
pipeline = transformers.pipeline(
|
| 145 |
+
"text-generation",
|
| 146 |
+
model=model,
|
| 147 |
+
torch_dtype=torch.float16,
|
| 148 |
+
device_map="auto",
|
| 149 |
+
)
|
| 150 |
+
|
| 151 |
+
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
|
| 152 |
+
print(outputs[0]["generated_text"])
|
| 153 |
+
```
|
| 154 |
+
|
| 155 |
+
|
| 156 |
+
## 🧩 Configuration
|
| 157 |
+
|
| 158 |
+
The model uses ChatML and should work well with this (as it is merged from models which (mostly) saw ChatML templates in training).
|
| 159 |
+
|
| 160 |
+
```yaml
|
| 161 |
+
models:
|
| 162 |
+
- model: mayflowergmbh/Wiedervereinigung-7b-dpo-laser
|
| 163 |
+
# no parameters necessary for base model
|
| 164 |
+
- model: flemmingmiguel/NeuDist-Ro-7B
|
| 165 |
+
parameters:
|
| 166 |
+
density: 0.60
|
| 167 |
+
weight: 0.30
|
| 168 |
+
- model: johannhartmann/Brezn3
|
| 169 |
+
parameters:
|
| 170 |
+
density: 0.65
|
| 171 |
+
weight: 0.40
|
| 172 |
+
- model: ResplendentAI/Flora_DPO_7B
|
| 173 |
+
parameters:
|
| 174 |
+
density: 0.6
|
| 175 |
+
weight: 0.3
|
| 176 |
+
merge_method: dare_ties
|
| 177 |
+
base_model: mayflowergmbh/Wiedervereinigung-7b-dpo-laser
|
| 178 |
+
parameters:
|
| 179 |
+
int8_mask: true
|
| 180 |
+
dtype: bfloat16
|
| 181 |
+
random_seed: 0
|
| 182 |
+
tokenizer_source: base
|
| 183 |
+
```
|