Maliktg7/qwen3-8b-opencodeinstruct-merged

qwen3-8b-opencodeinstruct

Qwen3-8B fine-tuned with QLoRA/SFT on a filtered, quality-scored slice of OpenCodeInstruct (CC BY 4.0), trained on 2x NVIDIA T4 GPUs (Kaggle) with PyTorch DDP via torchrun.

Results (pass@1)

Benchmark	Base Qwen3-8B	Fine-tuned
HumanEval+	<FILL IN from Phase 1>	<FILL IN from Phase 5>
MBPP+	<FILL IN from Phase 1>	<FILL IN from Phase 5>

Training details

• Base model: unsloth/Qwen3-8B-unsloth-bnb-4bit
• Method: QLoRA (r=16, alpha=32) + SFT
• Data: OpenCodeInstruct, filtered to average_test_score >= 0.8, exact-deduplicated
• Hardware: 2x Tesla T4 (Kaggle), DDP via torchrun
• Sequence length: 2048

Limitations

Trained on synthetic, LLM-generated instruction data; inherits any biases or gaps present in OpenCodeInstruct. Evaluated only on HumanEval+/MBPP+ — results may not generalize to other coding benchmarks or real-world repositories.