train.py

# ✅ Final train.py with JSONL input and EOS-as-pad (no stop_sequences)
import unsloth  # must be first
import pandas as pd
import torch
from datasets import Dataset
from transformers import TrainingArguments
from unsloth import FastLanguageModel
from trl import SFTTrainer
import os
import zipfile

# 1) Load Unsloth model + tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="HuggingFaceTB/SmolLM2-1.7B",
    max_seq_length=2048,
    dtype=torch.float16,
    load_in_4bit=True,
)

# 2) Reuse the existing eos_token as pad_token
eos = tokenizer.eos_token  # should be "<|endoftext|>"
tokenizer.pad_token = eos
model.config.pad_token_id = tokenizer.eos_token_id

# 3) Load & format your dataset from JSONL, always ending responses with EOS
#    Read the JSONL we generated (one JSON object per line with "instruction" & "response")
df = pd.read_json("data.jsonl", lines=True)
df["text"] = df.apply(
    lambda row: (
        "### Instruction:\n"
        + row["instruction"].strip()
        + "\n\n### Response:\n"
        + row["response"].strip()
        + eos
    ),
    axis=1
)
dataset = Dataset.from_pandas(df[["text"]])

# 4) Apply LoRA
model = FastLanguageModel.get_peft_model(
    model,
    r=8,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
)

# 5) Tokenize
def tokenize(example):
    return tokenizer(
        example["text"],
        truncation=True,
        padding="max_length",
        max_length=512,
    )

tokenized_dataset = dataset.map(tokenize, batched=True)

# 6) Training arguments
training_args = TrainingArguments(
    output_dir="./output_model",
    per_device_train_batch_size=2,
    num_train_epochs=3,
    learning_rate=2e-4,
    logging_steps=10,
    save_steps=100,
    fp16=True,
)

# 7) Initialize SFTTrainer (no stop_sequences here)
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    args=training_args,
    train_dataset=tokenized_dataset,
)

# 8) Train!
trainer.train()

# 9) Save the adapter
output_dir = "./output_model"
os.makedirs(output_dir, exist_ok=True)
model.save_pretrained(output_dir)

# 🔧 Zip for download
zip_path = "/home/user/app/model.zip"
try:
    with zipfile.ZipFile(zip_path, "w", zipfile.ZIP_DEFLATED) as z:
        for root, _, files in os.walk(output_dir):
            for fname in files:
                full = os.path.join(root, fname)
                rel = os.path.relpath(full, output_dir)
                z.write(full, rel)
    print(f"✅ Zipped model to {zip_path}")
except Exception as e:
    print(f"❌ Failed to zip model: {e}")