archit11/track_b_sft_merged

Track B SFT – Qwen2.5-Coder-1.5B (LoRA merged)

Fully merged fine-tuned model. LoRA weights have been merged into the base model weights — no PEFT library needed for inference.

Results

Metric	Baseline	Post-SFT	Δ
pass@1	0.565	0.804	+0.239
pass@3	0.783	0.848	+0.065

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("archit11/track_b_sft_merged", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("archit11/track_b_sft_merged")

prompt = "<|im_start|>user\nWrite a docstring for this function:\n```python\ndef add(a, b):\n    return a + b\n```<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0]))

Training Details

• Base: Qwen/Qwen2.5-Coder-1.5B
• Method: LoRA (r=16, alpha=32), merged
• Data: archit11/track_b_sft (257 train examples from verl corpus)
• Epochs: 3, LR: 2e-4, Hardware: T4 GPU (~56s training)