Use the following dataset to fine-tune llm_agents/tora-code-7b-v1.0 in order to improve the model's reasoning and planning abilities.
context window length: 16,384 prompt_type = "alpaca" max_tokens > 128 && < 16384
Total 177,333 samples 316 MB
- jondurbin/airoboros-2.2: Filter categories related to coding, reasoning and planning. 21,923 samples.
- Open-Orca/OpenOrca: Filter the 'cot' category in 1M GPT4 dataset. 62,973 samples.
- garage-bAInd/Open-Platypus: 100%, 22,760 samples.
- WizardLM/WizardLM_evol_instruct_V2_196k: Coding coversation part. 30,081 samples
- TokenBender/python_eval_instruct_51k: “python” in output .39,596 samples
50 samples/T=0.2/MaxTokens=512/Top_P=0.95
Code: https://github.com/uukuguy/speechless
How to Prompt the Model
This model accepts the Alpaca instruction format.
For example:
You are an intelligent programming assistant.
### Instruction:
Implement a linked list in C++
### Response:
HumanEval
| Metric | Value |
|---|---|
| humaneval-python | 52.44 |
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
MultiPL-E
| Metric | Value |
|---|---|
| python | 55.96 |
| java | 37.84 |
| javascript | 46.93 |
| cpp | 37.48 |
| rust | 29.01 |
| go | 28.99 |
| sh | 12.11 |
| julia | 31.47 |
| typescript | 47.80 |
LMEval
| Metric | Value |
|---|---|
| ARC | |
| HellaSwag | |
| MMLU | |
| TruthfulQA | |
| Average |
Parameters
| lr | 2e-4 |
| lr_scheduler_type | cosine |
| weight_decay | 0.0 |
| optim | paged_adamw_8bit |
| flash_attention | True |
| rerope | False |
| max_new_tokens | 16384 |
| num_train_epochs | 2 |
| bits | 4 |
| lora_r | 64 |
| lora_alpha | 256 |
| lora_dropout | 0.05 |
| double_quant | True |
| quant_type | nf4 |
| dataset_format | sharegpt |
| mini_batch_size | 2 |
| grandient_accumulation_steps | 32 |
| bf16 | True |
A100-40G x 4