DeepSeek-coder 微调训练记录

orange_の_呜

已于 2024-06-11 14:17:01 修改

阅读量5k

点赞数 10

文章标签： python 人工智能自然语言处理

于 2024-03-23 16:26:01 首次发布

本文链接：https://blog.csdn.net/m0_46268825/article/details/136937047

版权

简介

微调过程不再细说, 参考link进行即可. 主要是数据集.
在这里插入图片描述

1.3b模型微调训练占用资源信息

在这里插入图片描述
top信息

评估

根据DeepSeek-coder的Evaluation试进行对微调后的模型进行评估. 其中的评估库主要是evol-teacher和human-eval.

进入DeepSeek-coder下的Evaluation中的Human-eval文件夹, 新建一个eval_ins.sh文件, 填入以下内容

LANG="python"
OUPUT_DIR="output"
MODEL="deepseek-coder-1.3b-instruct"

CUDA_VISIBLE_DEVICES=0,1 python eval_instruct.py \
    --model "deepseek-ai/$MODEL" \  # 修改路径为微调保存的模型路径
    --output_path "$OUPUT_DIR/${LANG}.$MODEL.jsonl" \
    --language $LANG \
    --temp_dir $OUPUT_DIR

随后启动bash eval_ins.sh

model /home/LLM/DeepSeek-Coder/finetune/output/checkpoint-14500
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
load tokenizer <class 'transformers.models.llama.tokenization_llama_fast.LlamaTokenizerFast'> from /home/stlinpeiyang/lpy22/LLM/DeepSeek-Coder/finetune/output/checkpoint-14500 over.
Read 164 examples for evaluation over.
Generating:   1%|▌                                            | 2/164 [07:10<10:00:15, 222.32s/it]Failed to extract code block with error `list index out of range`:
>>> Task: Python/2
>>> Output:
def truncate_number(number: float) -> float:
    """ Given a positive floating point number, it can be decomposed into
    and integer part (largest integer smaller than given number) and decimals
    (leftover part always smaller than 1).

    Return the decimal part of the number.
    >>> truncate_number(3.5)
    0.5
    """
    integer_part = int(number)
    decimal_part = number - integer_part
    return decimal_part

Generating:   6%|██▋                                          | 10/164 [28:04<7:23:26, 172.77s/it]

耗时很久.

deepseek-coder-6.7b-instruct验证数据

(py_LLM) orange:~/lpy22/LLM/DeepSeek-Coder/Evaluation/HumanEval$ bash eval_6.7b.sh
model /home/LLM/text-generation-webui/models/deepseek-coder-6.7b-instruct
Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
load tokenizer <class 'transformers.models.llama.tokenization_llama_fast.LlamaTokenizerFast'> from /home/LLM/text-generation-webui/models/deepseek-coder-6.7b-instruct over.
Loading checkpoint shards: 100%|████████████████████████████████████████████| 2/2 [00:02<00:00,  1.12s/it]
Read 164 examples for evaluation over.
Generating: 100%|███████████████████████████████████████████████████| 164/164 [11:58:55<00:00, 263.02s/it]
Generate all over!!!
Save 164 processed examples into ./output/python.deepseek-coder-6.7b-instruct.jsonl over!
Reading samples...
100%|████████████████████████████████████████████████████████████████| 164/164 [00:00<00:00, 11619.16it/s]
Running test suites...
100%|███████████████████████████████████████████████████████████████████| 164/164 [00:14<00:00, 11.32it/s]
{'pass@1': 0.7865853658536586}
python {'pass@1': 0.7865853658536586} /home/LLM/text-generation-webui/models/deepseek-coder-6.7b-instruct

deepseek-coder-1.3b-instruct-finetune验证数据

Generating: 100%|███████████████████████████████████████████████████| 164/164 [10:58:16<00:00, 240.83s/it]
Generate all over!!!
Save 164 processed examples into ./output/python.deepseek-coder-1.3b-instruct-finetune.jsonl over!
Reading samples...
100%|███████████████████████████████████████████████████████████████████| 164/164 [00:00<00:00, 13670.37it
Running test suites...
100%|██████████████████████████████████████████████████████████████████████| 164/164 [00:13<00:00, 12.25it
{'pass@1': 0.6158536585365854}
python {'pass@1': 0.6158536585365854} /home/orange/LLM/DeepSeek-Coder/finetune/output/checkpoi14500

20140511
在这里插入图片描述