tiwanari/llm-jp-3-13b-it_lora

Uploaded model

• Developed by: tiwanari
• License: apache-2.0
• Finetuned from model : llm-jp/llm-jp-3-13b

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

jsonlファイルの出力方法

• 学習には　LLMのための日本語インストラクションデータ 公開ページのデータを活用。

  • ichikara-instruction-003-001-1.json を利用

• unslothを用いて、llm-jp/llm-jp-3-13bを4bit量子化のqLoRA設定でモデルを構築。ハイパーパラメータは下記のように設定。

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    dtype=None,
    load_in_4bit=True,
    trust_remote_code=True,
)

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 = ５１２,
)

• 訓練は以下のように設定した trl の Trainer によって行った。

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset=dataset["train"],
    max_seq_length = 512,
    dataset_text_field="formatted_text",
    packing = False,
    args = TrainingArguments(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 4,
        num_train_epochs = 1,
        logging_steps = 10,
        warmup_steps = 10,
        save_steps=100,
        save_total_limit=2,
        max_steps=-1,
        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",
    ),
)

• 上記を Hagging Face (このリポジトリ) にアップロード。
• 推論時には、このLoRAのアダプタをモデルに結合することで再現

model_id = "llm-jp/llm-jp-3-13b"
adapter_id = "tiwanari/llm-jp-3-13b-it_lora"

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_id,
    dtype=None,
    load_in_4bit=True,
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(model, adapter_id, token = HF_TOKEN)

• 結果はテストデータを用いて各行でのチェックを行って作成

datasets = []
with open("/content/elyza-tasks-100-TV_0.jsonl", "r") as f:
    item = ""
    for line in f:
      line = line.strip()
      item += line
      if item.endswith("}"):
        datasets.append(json.loads(item))
        item = ""

FastLanguageModel.for_inference(model)

results = []
for dt in tqdm(datasets):
  input = dt["input"]

  prompt = f"""### 指示\n{input}\n### 回答\n"""

  inputs = tokenizer([prompt], return_tensors = "pt").to(model.device)

  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]

  results.append({"task_id": dt["task_id"], "input": input, "output": prediction})

# 吐き出し
json_file_id = re.sub(".*/", "", adapter_id)
with open(f"/content/{json_file_id}_output.jsonl", 'w', encoding='utf-8') as f:
    for result in results:
        json.dump(result, f, ensure_ascii=False)
        f.write('\n')