code/text_classifier/bn/finetune/llama31_8b_32_3b.py

import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
import json
import ast
from unsloth import FastLanguageModel
import torch
from trl import SFTConfig, SFTTrainer
from datasets import Dataset
from unsloth.chat_templates import get_chat_template, standardize_sharegpt

# 1. Configuration
max_seq_length = 2048
dtype = None # Auto-detection
load_in_4bit = True
data_path = "/home/mshahidul/readctrl/data/finetuning_data/dataset_for_sft_support_check_list.json"
# model_name = "unsloth/Llama-3.1-8B"
model_name = "unsloth/Llama-3.2-3B-Instruct"
# 2. Load Model & Tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = model_name,
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
)

# 3. Add LoRA Adapters
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",
    random_state = 3407,
)

# 4. Data Prep (Conversation Format)
tokenizer = get_chat_template(tokenizer, chat_template="llama-3.1")

def formatting_prompts_func(examples):
    convos = examples["conversations"]
    texts = [
        tokenizer.apply_chat_template(
            convo,
            tokenize=False,
            add_generation_prompt=False,
        ).removeprefix("<bos>")
        for convo in convos
    ]
    return { "text" : texts, }

def parse_label_array(raw_text):
    text = (raw_text or "").strip()
    if not text:
        return []

    if "```" in text:
        text = text.replace("```json", "").replace("```", "").strip()

    start = text.find("[")
    end = text.rfind("]")
    if start != -1 and end != -1 and end > start:
        text = text[start : end + 1]

    parsed = None
    for parser in (json.loads, ast.literal_eval):
        try:
            parsed = parser(text)
            break
        except Exception:
            continue

    if not isinstance(parsed, list):
        return []

    normalized = []
    for item in parsed:
        if not isinstance(item, str):
            normalized.append("not_supported")
            continue
        label = item.strip().lower().replace("-", "_").replace(" ", "_")
        if label not in {"supported", "not_supported"}:
            label = "not_supported"
        normalized.append(label)
    return normalized

def extract_conversation_pair(conversations):
    user_prompt = ""
    gold_response = ""
    for message in conversations:
        role = message.get("role") or message.get("from")
        content = message.get("content", "")
        if role == "user" and not user_prompt:
            user_prompt = content
        elif role == "assistant" and not gold_response:
            gold_response = content
    return user_prompt, gold_response

def generate_prediction(user_prompt):
    prompt = tokenizer.apply_chat_template(
        [{"role": "user", "content": user_prompt}],
        tokenize=False,
        add_generation_prompt=True,
    )
    inputs = tokenizer([prompt], return_tensors="pt").to("cuda")
    with torch.inference_mode():
        outputs = model.generate(
            **inputs,
            max_new_tokens=128,
            do_sample=False,
            temperature=0.0,
            use_cache=True,
        )
    generated_tokens = outputs[0][inputs["input_ids"].shape[1]:]
    return tokenizer.decode(generated_tokens, skip_special_tokens=True).strip()

with open(data_path, "r", encoding="utf-8") as f:
    raw_data = json.load(f)

dataset = Dataset.from_list(raw_data)
dataset = standardize_sharegpt(dataset)
dataset = dataset.train_test_split(test_size=0.1, seed=3407, shuffle=True)

train_dataset = dataset["train"].map(formatting_prompts_func, batched=True)
test_dataset = dataset["test"]

# 5. Training
trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = train_dataset,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    packing = False,
    args = SFTConfig(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 4,
        warmup_steps = 5,
        max_steps = 60, # Increase for full training
        learning_rate = 2e-4,
        fp16 = not torch.cuda.is_bf16_supported(),
        bf16 = torch.cuda.is_bf16_supported(),
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = "outputs",
    ),
)
trainer.train()

# 6. Test-set Inference + Accuracy
FastLanguageModel.for_inference(model)
model.eval()
print("\n--- Testing Model on Test Set Samples ---")

for i in range(3):
    sample = test_dataset[i]
    user_prompt, _ = extract_conversation_pair(sample["conversations"])
    print(f"\nTest Sample {i+1} Prompt: {user_prompt}")
    decoded_output = generate_prediction(user_prompt)
    print(f"Model Response: {decoded_output}")

exact_match_correct = 0
label_correct = 0
label_total = 0
evaluated_samples = 0
parsed_prediction_count = 0

for sample in test_dataset:
    conversations = sample.get("conversations", [])
    user_prompt, gold_text = extract_conversation_pair(conversations)
    if not user_prompt:
        continue

    gold_labels = parse_label_array(gold_text)
    pred_text = generate_prediction(user_prompt)
    pred_labels = parse_label_array(pred_text)

    evaluated_samples += 1
    if pred_labels:
        parsed_prediction_count += 1

    if gold_labels and pred_labels == gold_labels:
        exact_match_correct += 1

    for pos, gold_label in enumerate(gold_labels):
        if pos < len(pred_labels) and pred_labels[pos] == gold_label:
            label_correct += 1
    label_total += len(gold_labels)

exact_match_accuracy = exact_match_correct / evaluated_samples if evaluated_samples else 0.0
label_accuracy = label_correct / label_total if label_total else 0.0

print("\n--- Test Accuracy ---")
print(f"Samples evaluated: {evaluated_samples}")
print(f"Parsed predictions: {parsed_prediction_count}")
print(f"Exact match accuracy: {exact_match_accuracy:.4f}")
print(f"Label accuracy: {label_accuracy:.4f}")
save_dir = f"/home/mshahidul/readctrl_model/support_checking_vllm/it_{model_name.split('/')[-1]}"
# 7. Save in FP16 Format (Merged)
# This creates a folder with the full model weights, not just adapters.
model.save_pretrained_merged(save_dir, tokenizer, save_method = "merged_16bit")
print(f"\nModel successfully saved in FP16 format to {save_dir}")