File size: 6,803 Bytes
5683186 3d6dd25 fbe2e0d 5683186 3d6dd25 5683186 3d6dd25 5683186 3d6dd25 5683186 3d6dd25 5683186 3d6dd25 5683186 3d6dd25 5683186 3d6dd25 e3fe64b 3d6dd25 5683186 04abfdf 5683186 3d6dd25 |
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 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 |
---
base_model: NousResearch/Llama-2-13b-hf
tags:
- llama-2
- instruct
- finetune
- alpaca
- gpt4
- synthetic data
- distillation
model-index:
- name: openhermes-13b
results: []
license: mit
language:
- en
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
# OpenHermes-13B
## Model description
OpenHermes 13B is the first fine tune of the Hermes dataset that has a fully open source dataset!
OpenHermes was trained on 242,000 entries of primarily GPT-4 generated data, from open datasets across the AI landscape, including:
- GPTeacher - General Instruct, Roleplay v1, Roleplay v2, and Code Instruct Datasets, by Teknium
- WizardLM (v1, evol_instruct 70k), by WizardLM Team/nlpxucan
- Airoboros GPT-4 (v1.0), by JonDurbin
- Camel-AI's domain expert datasets, by the Camel-AI Team
- CodeAlpaca, by Sahil2801
- GPT4-LLM and Unnatural Instructions, by Microsoft
Filtering included removal of OpenAI refusals, disclaimers, and "As an AI" type examples and more
The base dataset mix the model was trained on is identical to Nous-Hermes', minus the Nous-Instruct and PDACTL datasets which were private datasets.
The WANDB Project is public and can be examined at this link: https://wandb.ai/teknium1/openhermes/runs/openhermes-v2-fullft-13b
## Benchmark Information
## Benchmark Results
GPT-4All Benchmark Set
```
| Task |Version| Metric |Value | |Stderr|
|-------------|------:|--------|-----:|---|-----:|
|arc_challenge| 0|acc |0.5009|± |0.0146|
| | |acc_norm|0.5247|± |0.0146|
|arc_easy | 0|acc |0.8127|± |0.0080|
| | |acc_norm|0.7854|± |0.0084|
|boolq | 1|acc |0.8153|± |0.0068|
|hellaswag | 0|acc |0.6126|± |0.0049|
| | |acc_norm|0.7995|± |0.0040|
|openbookqa | 0|acc |0.3660|± |0.0216|
| | |acc_norm|0.4600|± |0.0223|
|piqa | 0|acc |0.7922|± |0.0095|
| | |acc_norm|0.8112|± |0.0091|
|winogrande | 0|acc |0.7293|± |0.0125|
```
AGI-Eval
```
| Task |Version| Metric |Value | |Stderr|
|------------------------------|------:|--------|-----:|---|-----:|
|agieval_aqua_rat | 0|acc |0.2008|± |0.0252|
| | |acc_norm|0.2126|± |0.0257|
|agieval_logiqa_en | 0|acc |0.3410|± |0.0186|
| | |acc_norm|0.3564|± |0.0188|
|agieval_lsat_ar | 0|acc |0.2261|± |0.0276|
| | |acc_norm|0.2174|± |0.0273|
|agieval_lsat_lr | 0|acc |0.3725|± |0.0214|
| | |acc_norm|0.3373|± |0.0210|
|agieval_lsat_rc | 0|acc |0.4684|± |0.0305|
| | |acc_norm|0.4572|± |0.0304|
|agieval_sat_en | 0|acc |0.6553|± |0.0332|
| | |acc_norm|0.5971|± |0.0343|
|agieval_sat_en_without_passage| 0|acc |0.4515|± |0.0348|
| | |acc_norm|0.4029|± |0.0343|
|agieval_sat_math | 0|acc |0.3273|± |0.0317|
| | |acc_norm|0.2636|± |0.0298|
```
BigBench Reasoning Test
```
| Task |Version| Metric |Value | |Stderr|
|------------------------------------------------|------:|---------------------|-----:|---|-----:|
|bigbench_causal_judgement | 0|multiple_choice_grade|0.5368|± |0.0363|
|bigbench_date_understanding | 0|multiple_choice_grade|0.7127|± |0.0236|
|bigbench_disambiguation_qa | 0|multiple_choice_grade|0.3023|± |0.0286|
|bigbench_geometric_shapes | 0|multiple_choice_grade|0.1003|± |0.0159|
| | |exact_str_match |0.0000|± |0.0000|
|bigbench_logical_deduction_five_objects | 0|multiple_choice_grade|0.2720|± |0.0199|
|bigbench_logical_deduction_seven_objects | 0|multiple_choice_grade|0.1986|± |0.0151|
|bigbench_logical_deduction_three_objects | 0|multiple_choice_grade|0.4500|± |0.0288|
|bigbench_movie_recommendation | 0|multiple_choice_grade|0.2880|± |0.0203|
|bigbench_navigate | 0|multiple_choice_grade|0.5000|± |0.0158|
|bigbench_reasoning_about_colored_objects | 0|multiple_choice_grade|0.5390|± |0.0111|
|bigbench_ruin_names | 0|multiple_choice_grade|0.3906|± |0.0231|
|bigbench_salient_translation_error_detection | 0|multiple_choice_grade|0.1844|± |0.0123|
|bigbench_snarks | 0|multiple_choice_grade|0.5249|± |0.0372|
|bigbench_sports_understanding | 0|multiple_choice_grade|0.5335|± |0.0159|
|bigbench_temporal_sequences | 0|multiple_choice_grade|0.2980|± |0.0145|
|bigbench_tracking_shuffled_objects_five_objects | 0|multiple_choice_grade|0.2048|± |0.0114|
|bigbench_tracking_shuffled_objects_seven_objects| 0|multiple_choice_grade|0.1297|± |0.0080|
|bigbench_tracking_shuffled_objects_three_objects| 0|multiple_choice_grade|0.4500|± |0.0288|
```
This is a slight improvement on GPT4ALL Suite and BigBench Suite, with a degredation in AGIEval compared to the original hermes.
Average Score Comparison between Nous-Hermes Llama-2 and OpenHermes Llama-2:
```
| Bench | Nous-Hermes | OpenHermes | Change |
|------------------------------|------------:|------------|--------|
|GPT4All | 70.00| 70.36| +0.36|
|------------------------------------------------------------------|
|BigBench | 36.57| 36.75| +0.18|
|------------------------------------------------------------------|
|AGI Eval | 37.20| 35.56| -1.64|
```
## Training procedure
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 2
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- gradient_accumulation_steps: 8
- total_train_batch_size: 128
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 300
- num_epochs: 3
### Framework versions
- Transformers 4.34.0.dev0
- Pytorch 2.0.1+cu118
- Datasets 2.14.4
- Tokenizers 0.13.3 |