Framework: Unsloth + TRL SFTTrainer
Launch: 2-GPU DDP via torchrun --nproc_per_node=2
GPUs: x2 H200
LoRA rank: 64
LoRA alpha: 128
Max sequence length: 2048
Epochs configured: 5
Early stopping: metric-based, not epoch-based
Metric early-stop rule: stop when combined GSM8K/MATH-500 probe score does not improve by at least 0.75 percentage points for 2 checks

Dataset mixture

70% nvidia/OpenMathInstruct-2
15% meta-math/MetaMathQA
10% TIGER-Lab/MathInstruct
5% AI-MO/NuminaMath-CoT

Exact decontamination was applied against GSM8K test and MATH-500 test question text hashes.

Final evaluation

Benchmark	Accuracy
GSM8K test	50.57%
MATH-500 test	28.80%
Combined average	39.68%

Usage

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Kiffaz11/llama-3.2-1b-instruct-math-sft-lora",
    max_seq_length=4096,
    dtype=None,
    load_in_4bit=False,
)

FastLanguageModel.for_inference(model)

messages = [
    {
        "role": "user",
        "content": "Solve step by step. Put the final answer in \\boxed{}. What is 17*23?",
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    do_sample=False,
)

print(
    tokenizer.decode(
        outputs[0][inputs["input_ids"].shape[-1]:],
        skip_special_tokens=True,
    )
)