llm-judge-full-backup

Full training folder backup - Toàn bộ checkpoints và models.

📂 Cấu trúc Folder

train_llm_judge_v2/
├── checkpoint-150/          # Checkpoint tại step 150
├── checkpoint-200/          # Checkpoint tại step 200
├── checkpoint-210/          # Checkpoint tại step 210
├── final_model/             # Model cuối cùng (merged)
├── lora_adapters/           # LoRA adapters
├── README.md
├── zero_shot_metrics.json
└── zero_shot_results.csv

🚀 Sử Dụng

1️⃣ Clone Repo

git lfs install
git clone https://huggingface.co/ImNotTam/llm-judge-full-backup
cd llm-judge-full-backup

2️⃣ Load LoRA Adapters (Nhẹ nhất - khuyến nghị)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="ImNotTam/llm-judge-full-backup",
    subfolder="lora_adapters",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

# Enable inference mode
FastLanguageModel.for_inference(model)

# Test
prompt = "Đánh giá response này..."
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

3️⃣ Load Final Model

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "ImNotTam/llm-judge-full-backup",
    subfolder="final_model",
    device_map="auto",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("ImNotTam/llm-judge-full-backup", subfolder="final_model")

# Inference
inputs = tokenizer("Your prompt", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0]))

4️⃣ Resume Training từ Checkpoint

from transformers import Trainer, TrainingArguments

# Load checkpoint muốn resume
model = AutoModelForCausalLM.from_pretrained(
    "ImNotTam/llm-judge-full-backup",
    subfolder="checkpoint-210",  # Chọn checkpoint
    device_map="auto"
)

# Continue training
trainer = Trainer(
    model=model,
    args=TrainingArguments(
        output_dir="./continue_training",
        # ... your training args
    ),
)
trainer.train(resume_from_checkpoint=True)

5️⃣ Fine-tune Tiếp từ LoRA Adapter

from unsloth import FastLanguageModel
from trl import SFTTrainer

# Load LoRA adapter
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="ImNotTam/llm-judge-full-backup",
    subfolder="lora_adapters",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

# Add LoRA config để train tiếp
model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
    lora_alpha=16,
    lora_dropout=0,
    bias="none",
    use_gradient_checkpointing="unsloth",
)

# Train với data mới
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=your_new_dataset,
    # ... training args
)
trainer.train()

6️⃣ Xem Metrics và Results

import json
import pandas as pd

# Load metrics
with open("zero_shot_metrics.json", "r") as f:
    metrics = json.load(f)
print("📊 Metrics:", metrics)

# Load results
results = pd.read_csv("zero_shot_results.csv")
print("\n📈 Results:")
print(results.head())

📋 Nội Dung Repo

Folder/File	Mô tả	Kích thước
lora_adapters/	LoRA adapters (nhẹ)	~50-100 MB
final_model/	Model merged đầy đủ	~4-8 GB
checkpoint-150/	Training checkpoint	~4-8 GB
checkpoint-200/	Training checkpoint	~4-8 GB
checkpoint-210/	Training checkpoint	~4-8 GB
zero_shot_metrics.json	Evaluation metrics	<1 MB
zero_shot_results.csv	Detailed results	<1 MB

💡 Khuyến Nghị

• Inference nhanh: Dùng lora_adapters/
• Production: Dùng final_model/
• Train tiếp: Load lora_adapters/ + add LoRA config
• Resume training: Load checkpoint cụ thể

📦 Requirements

pip install unsloth transformers torch trl

📄 License

Apache 2.0