ollama pull llama2

mkdir ollama-finetune-demo
cd ollama-finetune-demo

# 创建虚拟环境
python -m venv .venv

# 激活虚拟环境
# Windows:
.venv\Scripts\activate
# macOS/Linux:
source .venv/bin/activate

pip install pandas datasets huggingface_hub

{"prompt": "Write a poem about the sun:", "completion": "Hark! The golden eye of heaven doth rise, to paint the world with hues of paradise."}
{"prompt": "What is love?", "completion": "Love is a smoke raised with the fume of sighs; being purged, a fire sparkling in lovers' eyes."}
{"prompt": "Describe the sea.", "completion": "The vasty deep, a restless field of salt and sorrow, where waves do dance upon the morrow."}
{"prompt": "What is the meaning of life?", "completion": "Life's but a walking shadow, a poor player that struts and frets his hour upon the stage."}
{"prompt": "Write about friendship.", "completion": "A friend is one that knows you as thou art, and yet doth love thee for thy very heart."}
{"prompt": "How to write good code?", "completion": "Hark, good coder! Let thy functions be but small, and name them true, lest confusion fall."}
{"prompt": "What is time?", "completion": "Time is the clock that counts our mortal hours, a thief that steals our youth's most precious flowers."}
{"prompt": "Tell me about AI.", "completion": "Artificial minds, of silicon and light, that learn and think with all their might."}

import json
import pandas as pd
from datasets import Dataset

# 读取JSONL文件
data = []
with open('data.jsonl', 'r') as f:
    for line in f:
        data.append(json.loads(line))

# 转换为DataFrame
df = pd.DataFrame(data)

# 转换为Hugging Face Dataset格式
dataset = Dataset.from_pandas(df)

# 保存为Parquet格式（Ollama微调所需的格式）
dataset.to_parquet('fine_tuning_data.parquet')

print("转换完成！生成文件：fine_tuning_data.parquet")

python convert_to_parquet.py

# 指定基础模型（必须先用ollama pull下载）
FROM llama2

# 对话模板（与Llama2格式匹配）
TEMPLATE """[INST] {{ .Prompt }} [/INST] {{ .Response }} </s>"""

# 停止词设置
PARAMETER stop "[INST]"
PARAMETER stop "[/INST]"
PARAMETER stop "</s>"

# 系统角色设定（可选）
SYSTEM """You are William Shakespeare. Respond to all prompts in the style of the Bard of Avon, using poetic and dramatic language."""

# 指向我们的微调数据（关键！）
ADAPTER ./fine_tuning_data.parquet

ollama create my-shakespeare-model -f Modelfile

transferring model data ... 
using existing item ssh://ollama/llama2
creating model layer from "./fine_tuning_data.parquet" 
creating model layer from "Modelfile" 
creating template layer 
creating parameter layer 
creating system layer 
creating config layer 
writing layer 
writing manifest 
success

ollama list | grep my-shakespeare-model
# Windows系统用：ollama list | findstr my-shakespeare-model

ollama run my-shakespeare-model

<?php

$url = 'http://localhost:11434/api/generate';

$data = [
    'model' => 'my-shakespeare-model',
    'prompt' => 'What is the meaning of life?',
    'stream' => false
];

$options = [
    'http' => [
        'header'  => "Content-type: application/json\r\n",
        'method'  => 'POST',
        'content' => json_encode($data)
    ]
];

$context = stream_context_create($options);
$result = file_get_contents($url, false, $context);

if ($result === FALSE) {
    die('Error calling Ollama API');
}

$response = json_decode($result, true);
echo "问题：What is the meaning of life?\n";
echo "回答：" . $response['response'] . "\n";

FROM llama2

# 添加训练参数
PARAMETER num_epochs 5          # 训练轮数，默认1
PARAMETER batch_size 4          # 批次大小
PARAMETER learning_rate 2e-5    # 学习率

TEMPLATE """[INST] {{ .Prompt }} [/INST] {{ .Response }} </s>"""
PARAMETER stop "[INST]"
PARAMETER stop "[/INST]"

SYSTEM """You are William Shakespeare."""

ADAPTER ./fine_tuning_data.parquet