train.py

"""
Fine-tuning Qwen2.5-3B-Instruct for football prediction extraction
Fixes from original: target_modules, validation split, scheduler, checkpoint saving
"""

from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, BitsAndBytesConfig
from peft import LoraConfig
from trl import SFTTrainer, DataCollatorForCompletionOnlyLM
import torch

# ─────────────────────────────────────────────
# CONFIG
# ─────────────────────────────────────────────
MODEL_NAME   = "Qwen/Qwen2.5-3B-Instruct"
OUTPUT_DIR   = "./football-extractor"
TRAIN_FILE   = "train_dataset.jsonl"
VAL_FILE     = "val_dataset.jsonl"

# ─────────────────────────────────────────────
# LOAD DATA
# ─────────────────────────────────────────────
dataset = load_dataset("json", data_files={"train": TRAIN_FILE, "validation": VAL_FILE})
print(f"Train: {len(dataset['train'])} | Val: {len(dataset['validation'])}")

# ─────────────────────────────────────────────
# TOKENIZER
# ─────────────────────────────────────────────
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"   # important for causal LM training

# ─────────────────────────────────────────────
# QUANTIZATION (4-bit QLoRA)
# ─────────────────────────────────────────────
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,   # bfloat16 is more stable than float16
    bnb_4bit_use_double_quant=True,           # saves a bit more VRAM
)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    quantization_config=bnb_config,
    device_map="auto",
    attn_implementation="eager",              # avoids flash-attn issues on Colab
)
model.config.use_cache = False               # required for gradient checkpointing

# ─────────────────────────────────────────────
# LORA CONFIG
# ─────────────────────────────────────────────
lora_config = LoraConfig(
    r=8,                     # smaller r is fine for simple extraction
    lora_alpha=16,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    # Explicitly target attention + MLP layers for Qwen2.5
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj"
    ],
)

# ─────────────────────────────────────────────
# FORMAT FUNCTION
# ─────────────────────────────────────────────
def format_example(example):
    """Apply Qwen2.5 chat template to each training example."""
    return tokenizer.apply_chat_template(
        example["messages"],
        tokenize=False,
        add_generation_prompt=False
    )

# ─────────────────────────────────────────────
# TRAINING ARGS
# ─────────────────────────────────────────────
training_args = TrainingArguments(
    output_dir=OUTPUT_DIR,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4,       # effective batch = 4
    gradient_checkpointing=True,         # saves VRAM
    learning_rate=2e-4,
    num_train_epochs=3,
    lr_scheduler_type="cosine",          # smooth decay
    warmup_ratio=0.05,                   # 5% warmup steps
    logging_steps=10,
    eval_strategy="epoch",               # evaluate after each epoch
    save_strategy="epoch",               # save checkpoint each epoch
    save_total_limit=2,                  # keep only last 2 checkpoints
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss",
    fp16=False,
    bf16=True,                           # use bfloat16 if your GPU supports it
    report_to="none",                    # set to "wandb" if you want tracking
)

# ─────────────────────────────────────────────
# TRAINER
# ─────────────────────────────────────────────
trainer = SFTTrainer(
    model=model,
    train_dataset=dataset["train"],
    eval_dataset=dataset["validation"],
    peft_config=lora_config,
    args=training_args,
    formatting_func=format_example,
    max_seq_length=512,                  # extraction tasks are short
)

trainer.train()
trainer.save_model(OUTPUT_DIR)
tokenizer.save_pretrained(OUTPUT_DIR)
print(f"✅ Adapter saved to {OUTPUT_DIR}")