TIMPS-Coder-7B

TIMPS-Coder-7B is a code-generation model built by fine-tuning Qwen2.5-Coder-7B-Instruct through a 3-step pipeline: SFT → GRPO → DPO, achieving state-of-the-art results on HumanEval (98.8% pass@1) among 7B-9B code models.

---

Benchmark Results

Benchmark	TIMPS-Coder-7B	Qwen2.5-Coder-7B-Instruct	Delta
HumanEval pass@1	98.8%	86.6%	+12.2pp
HumanEval+ pass@1	82.9%	71.3%	+11.6pp
MBPP+ pass@1	73.3%	69.6%	+3.7pp

---

Full Comparison Data

Model	HumanEval	HumanEval+	MBPP	MBPP+	Params
TIMPS-Coder-7B	98.8	82.9	5.4	73.3	7B
Qwen2.5-Coder-7B-Instruct	86.6	71.3	82.0	69.6	7.6B
Qwen2.5-Coder-7B	89.6	76.2	84.0	72.0	7.6B
DeepSeek-Coder-7B-Instruct-v1.5	84.1	70.8	79.6	68.4	7.1B
CodeLlama-7B-Instruct	53.7	44.5	55.6	45.0	6.7B
CodeGemma-7B-it	56.1	46.9	61.8	50.6	7.0B
StarCoder2-7B	40.2	32.9	46.0	36.5	7.0B
Llama-3.1-8B-Instruct	72.6	61.0	70.8	58.7	8.0B
Phi-3.5-mini-instruct	68.8	57.9	73.0	61.3	3.8B
Gemma-2-9B-it	54.3	44.5	59.6	49.3	9.2B

Note: MBPP pass@1 (5.4%) for TIMPS-Coder-7B is notably low; this is a known evaluation artifact. The extended MBPP+ score (73.3%) is the more reliable indicator of MBPP performance.

---

Model Details

• Base Model: Qwen/Qwen2.5-Coder-7B-Instruct
• Training Pipeline: SFT → GRPO → DPO (LoRA adapters merged via PEFT)
• Parameters: 7B (8B total)
• Precision: BF16
• License: Apache-2.0
• HuggingFace: sandeeprdy1729/TIMPS-Coder-7B

---

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "sandeeprdy1729/TIMPS-Coder-7B",
    device_map="auto",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("sandeeprdy1729/TIMPS-Coder-7B")

messages = [{"role": "user", "content": "Write a fibonacci function."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(inputs, max_new_tokens=512)[0]))

---

Citation

@misc{timps-coder-7b,
  title={TIMPS-Coder-7B: SFT + GRPO + DPO Fine-tuned Code Model},
  author={sandeeprdy1729},
  year={2026},
  url={https://huggingface.co/sandeeprdy1729/TIMPS-Coder-7B}
}

---

License

This project is licensed under the Apache-2.0 License.