Update README.md

README.md

@@ -20,3 +20,232 @@ language:
 This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
 
 [<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
+
+!pip uninstall unsloth -y
+!pip install --upgrade --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
+!pip install --upgrade torch
+!pip install --upgrade xformers
+
+# Install Flash Attention 2 for softcapping support
+import torch
+if torch.cuda.get_device_capability()[0] >= 8:
+    !pip install --no-deps packaging ninja einops "flash-attn>=2.6.3"
+
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+from unsloth import FastLanguageModel
+import torch
+
+max_seq_length = 512  
+dtype = None  
+load_in_4bit = True  
+
+model_id = "llm-jp/llm-jp-3-13b"
+new_model_id = "llm-jp-3-13b-last"  
+
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name=model_id,
+    dtype=dtype,
+    load_in_4bit=load_in_4bit,
+    trust_remote_code=True,
+)
+
+# SFT用のモデルを用意
+model = FastLanguageModel.get_peft_model(
+    model,
+    r=32,
+    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
+                    "gate_proj", "up_proj", "down_proj"],
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    use_gradient_checkpointing="unsloth",
+    random_state=3407,
+    use_rslora=False,
+    loftq_config=None,
+    max_seq_length=max_seq_length,
+)
+
+# https://huggingface.co/settings/tokens
+HF_TOKEN = "your-token"  # @param {type:"string"}
+
+from datasets import load_dataset, concatenate_datasets
+
+# データセットのロード
+ichikara_dataset = load_dataset("json", data_files="/content/ichikara-instruction-003-001-1.json")
+elyza_dataset = load_dataset("elyza/ELYZA-tasks-100")
+
+EOS_TOKEN = tokenizer.eos_token  # 
+
+# 学習時のプロンプトフォーマットの定義
+prompt = """### 指示
+{}
+### 回答
+{}"""
+
+"""
+formatting_prompts_func: 各データをプロンプトに合わせた形式に合わせる
+"""
+def formatting_prompts_func(examples):
+    input = examples["text"]  
+    output = examples["output"]  
+    text = prompt.format(input, output) + EOS_TOKEN  
+    return {"formatted_text": text} 
+
+# ichikara-instruction のデータフォーマット
+ichikara_dataset = ichikara_dataset.map(
+    formatting_prompts_func,
+    num_proc=4,  
+)
+
+# ELYZA-tasks-100 データセットのフォーマット関数
+def elyza_formatting_prompts_func(examples):
+    input = examples["input"]  
+    output = examples["output"]  
+    text = prompt.format(input, output) + EOS_TOKEN  
+    return {"formatted_text": text}
+
+# ELYZA-tasks-100 のデータフォーマット
+elyza_dataset = elyza_dataset.map(
+    elyza_formatting_prompts_func,
+    num_proc=4  
+)
+
+from datasets import concatenate_datasets
+
+# ichikara-instruction と ELYZA-tasks-100 を統合
+combined_dataset = concatenate_datasets([
+    ichikara_dataset["train"],  
+    elyza_dataset["test"]      
+])
+
+# データ品質チェック
+# 1. ランダムサンプルを確認
+import random
+sample_indices = random.sample(range(len(combined_dataset)), 10)
+for idx in sample_indices:
+    print(combined_dataset[idx]["formatted_text"])
+
+# 2. 自動検査ルール
+# 短すぎるデータをチェック（Noneチェックを追加）
+short_data = combined_dataset.filter(
+    lambda x: x["input"] is not None and x["output"] is not None and (len(x["input"]) < 5 or len(x["output"]) < 5)
+)
+print(f"\n短すぎるデータ数: {len(short_data)}")
+
+# 指示と回答が同一のデータ（Noneチェックを追加）
+duplicate_data = combined_dataset.filter(
+    lambda x: x["input"] is not None and x["output"] is not None and x["input"].strip() == x["output"].strip()
+)
+print(f"\n指示と回答が同一のデータ数: {len(duplicate_data)}")
+
+# 問題のあるデータをフィルタリング（Noneチェックを追加）
+filtered_dataset = combined_dataset.filter(
+    lambda x: x["input"] is not None and x["output"] is not None and len(x["input"]) > 5 and len(x["output"]) > 5 and x["input"].strip() != x["output"].strip()
+)
+
+print(f"元のデータ数: {len(combined_dataset)}")
+print(f"フィルタリング後のデータ数: {len(filtered_dataset)}")
+print(f"除外されたデータ数: {len(combined_dataset) - len(filtered_dataset)}")
+
+# フィルタリング後のデータの例を確認
+print(filtered_dataset[0])
+
+"""
+training_arguments: 学習の設定
+"""
+from trl import SFTTrainer
+from transformers import TrainingArguments
+from unsloth import is_bfloat16_supported
+
+trainer = SFTTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    train_dataset=filtered_dataset,
+    max_seq_length=max_seq_length,
+    dataset_text_field="formatted_text",
+    packing=False,
+    args=TrainingArguments(
+        per_device_train_batch_size=2,
+        gradient_accumulation_steps=4,
+        num_train_epochs=3,
+        logging_steps=10,
+        warmup_steps=10,
+        save_steps=50,
+        save_total_limit=2,
+        max_steps=200,
+        learning_rate=2e-4,
+        fp16=not is_bfloat16_supported(),
+        bf16=is_bfloat16_supported(),
+        group_by_length=True,
+        seed=3407,
+        output_dir="outputs",
+        report_to="none",
+    ),
+)
+
+#@title 学習実行
+trainer_stats = trainer.train()
+
+import json
+from datasets import load_dataset
+
+dataset = load_dataset("json", data_files="/content/elyza-tasks-100-TV_0.jsonl", split="train")
+
+datasets = []  
+with open("/content/elyza-tasks-100-TV_0.jsonl", "r", encoding="utf-8") as f:  
+    item = ""
+    for line in f:
+        line = line.strip()
+        item += line
+        if item.endswith("}"):
+            datasets.append(json.loads(item))
+            item = ""
+
+from tqdm import tqdm
+import json
+
+# 推論するためにモデルのモードを変更
+FastLanguageModel.for_inference(model)
+
+results = []
+for dt in tqdm(datasets):
+    try:
+        input_text = dt["input"]
+
+        # プロンプトを生成
+        prompt = f"### 指示\n{input_text}\n次の要件を満たしてください：\n1. 簡潔に回答する。\n2. 必要なら箇条書きを使用して要点を整理する。\n3. 指示された内容に忠実に答える。\n### 回答\n"
+
+
+        # トークナイズ
+        inputs = tokenizer([prompt], return_tensors="pt").to(model.device)
+
+        # 推論
+        outputs = model.generate(
+            **inputs,
+            max_new_tokens=512,
+            use_cache=True,
+            do_sample=False,
+            repetition_penalty=1.2,
+        )
+        prediction = tokenizer.decode(outputs[0], skip_special_tokens=True).split('\n### 回答')[-1]
+
+        # 結果を保存
+        results.append({"task_id": dt["task_id"], "input": input_text, "output": prediction})
+    except Exception as e:
+        print(f"Error processing task_id {dt.get('task_id', 'Unknown')}: {e}")
+        results.append({"task_id": dt.get("task_id", "Unknown"), "input": dt.get("input", ""), "output": "Error"})
+
+
+# 結果をJSONL形式で保存
+output_file_jsonl = "/content/llm-jp-3-13b-last.jsonl"
+with open(output_file_jsonl, "w", encoding="utf-8") as f:
+    for result in results:
+        f.write(json.dumps(result, ensure_ascii=False) + "\n")
+
+model.push_to_hub_merged(
+    new_model_id,
+    tokenizer=tokenizer,
+    save_method="lora",
+    token=HF_TOKEN,
+    private=True
+)