Use the following dataset to fine-tune Open-Orca/Mistral-7B-OpenOrca in order to improve the model's reasoning and planning abilities.
Total 201,981 samples.
- jondurbin/airoboros-2.2: Filter categories related to coding, reasoning and planning. 23,462 samples.
- Open-Orca/OpenOrca: Filter the 'cot' category in 1M GPT4 dataset. 74,440 samples.
- garage-bAInd/Open-Platypus: 100%, 24,926 samples.
- WizardLM/WizardLM_evol_instruct_V2_196k: Coding coversation part. 30,185 samples
- TokenBender/python_eval_instruct_51k: “python” in output .40,309 samples
- Spider: 8,659 samples
Code: https://github.com/uukuguy/speechless
HumanEval
| Metric | Value |
|---|---|
| humaneval-python | 47.561 |
CodeLlama-34B-Python: 53.29
CodeLlama-34B-Instruct: 50.79
CodeLlama-13B-Instruct: 50.6
CodeLlama-34B: 45.11
CodeLlama-13B-Python: 42.89
CodeLlama-13B: 35.07
lm-evaluation-harness
| Metric | Value |
|---|---|
| ARC | 59.64 |
| HellaSwag | 82.25 |
| MMLU | 61.33 |
| TruthfulQA | 48.45 |
| Average | 62.92 |
Parameters
| lr | 2e-4 |
| lr_scheduler_type | cosine |
| weight_decay | 0.0 |
| optim | paged_adamw_8bit |
| flash_attention | True |
| rerope | False |
| max_new_tokens | 4096 |
| num_train_epochs | 2 |
| bits | 4 |
| lora_r | 64 |
| lora_alpha | 16 |
| lora_dropout | 0.05 |
| double_quant | True |
| quant_type | nf4 |
| dataset_format | airoboros |
| mini_batch_size | 2 |
| grandient_accumulation_steps | 32 |
| bf16 | True |
A100-40G x 4
| epoch | 2.0 |
| etrain_loss | 0.4708 |
| etrain_runtime | 12:12:53.64 |
| etrain_samples_per_second | 9.002 |
| etrain_steps_per_second | 0.07 |
| eeval_loss | 0.4851 |
| eeval_runtime | 0:00:10.31 |
| eeval_samples_per_second | 19.385 |
| eeval_steps_per_second | 4.846 |
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 55.33 |
| ARC (25-shot) | 59.64 |
| HellaSwag (10-shot) | 82.25 |
| MMLU (5-shot) | 61.33 |
| TruthfulQA (0-shot) | 48.45 |
| Winogrande (5-shot) | 77.51 |
| GSM8K (5-shot) | 8.26 |
| DROP (3-shot) | 49.89 |