train_mini_test.py

"""
Quick smoke-test script to verify your environment works before full training.
Runs on just 20 training + 5 eval samples with 10 steps.
"""
import torch
from datasets import load_dataset
from unsloth import FastLanguageModel
from trl import SFTConfig, SFTTrainer
from transformers import AutoTokenizer, set_seed

set_seed(3407)

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-2B", trust_remote_code=True)

def preprocess(example):
    text = tokenizer.apply_chat_template(
        example["messages"], tokenize=False, add_generation_prompt=False
    )
    return {"text": text}

print("Loading dataset...")
train_dataset = load_dataset("OmAlve/reading-steiner-data", split="train[:20]")
eval_dataset = load_dataset("OmAlve/reading-steiner-data", split="eval[:5]")

train_dataset = train_dataset.map(preprocess, remove_columns=["messages"])
eval_dataset  = eval_dataset.map(preprocess,  remove_columns=["messages"])

print("Loading model with Unsloth...")
model, _ = FastLanguageModel.from_pretrained(
    model_name="Qwen/Qwen3.5-2B",
    max_seq_length=4096,
    dtype=torch.float16,
    load_in_4bit=True,
)

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=True,
    random_state=3407,
)

args = SFTConfig(
    output_dir="./mini_out",
    max_length=4096,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4,
    max_steps=10,
    learning_rate=2e-4,
    warmup_ratio=0.05,
    lr_scheduler_type="cosine",
    logging_steps=2,
    logging_first_step=True,
    eval_strategy="steps",
    eval_steps=5,
    save_strategy="no",
    bf16=torch.cuda.is_bf16_supported(),
    fp16=not torch.cuda.is_bf16_supported(),
    gradient_checkpointing=True,
    disable_tqdm=True,
    report_to=["trackio"],
    seed=3407,
    push_to_hub=False,
)

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    args=args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
)

print("Starting mini training run...")
trainer.train()
print("Mini test passed! Your environment is ready for full training.")