search

bigcode-project / bigcode-evaluation-harness

A framework for the evaluation of autoregressive code generation language models.
Apache License 2.0
702 stars 180 forks source link

support left padding and prefix post-processing for models like chatglm #184

Closed loubnabnl closed 5 months ago

loubnabnl commented 5 months ago

fixes https://github.com/bigcode-project/bigcode-evaluation-harness/issues/182 ChatGLM3-6B requires left padding which wasn't supported in the harness, you need to add --left_padding flag to force it. 

accelerate launch main.py \
    --model THUDM/chatglm3-6b \
    --tasks humaneval \
    --n_samples 1 \
    --batch_size 1 \
    --allow_code_execution \
    --use_auth_token \
    --do_sample False \
    --precision fp16 \
    --trust_remote_code \
    --save_generations \
    --left_padding

output

{
  "humaneval": {
    "pass@1": 0.29878048780487804
  },
  "config": {
    "prefix": "",
    "do_sample": false,
    "temperature": 0.2,
    "top_k": 0,
    "top_p": 0.95,
    "n_samples": 1,
    "eos": "<|endoftext|>",
    "seed": 0,
    "model": "THUDM/chatglm3-6b",
    "modeltype": "causal",
    "peft_model": null,
    "revision": null,
    "use_auth_token": true,
    "trust_remote_code": true,
    "tasks": "humaneval",
    "instruction_tokens": null,
    "batch_size": 1,
    "max_length_generation": 512,
    "precision": "fp16",
    "load_in_8bit": false,
    "load_in_4bit": false,
    "left_padding": true,
    "limit": null,
    "limit_start": 0,
    "save_every_k_tasks": -1,
    "postprocess": true,
    "allow_code_execution": true,
    "generation_only": false,
    "load_generations_path": null,
    "load_data_path": null,
    "metric_output_path": "evaluation_results.json",
    "save_generations": true,
    "load_generations_intermediate_paths": null,
    "save_generations_path": "generations.json",
    "save_references": false,
    "save_references_path": "references.json",
    "prompt": "prompt",
    "max_memory_per_gpu": null,
    "check_references": false
  }
}