TIMPS-Coder-0.5B

TIMPS-Coder v3 — Elite Bug-Fixing Assistant (0.5B)

A 0.5B parameter coding model fine-tuned to think before it codes — specialising in bug analysis, code review, algorithm problem-solving, and agentic planning.
Built by Sandeep Reddy · TIMPS · Made in India 🇮🇳

Model Summary

Benchmark Results — 25 Tests, 5 Dimensions

Evaluated on 3_benchmark_ollama.py.
Scoring: 2 pts = complete correct answer with code · 1 pt = partial · 0 = wrong/refused.

Quick Start

Ollama (recommended)

ollama pull sandeeprdy1729/timps-coder
ollama run sandeeprdy1729/timps-coder

Python (Transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer

model     = AutoModelForCausalLM.from_pretrained("sandeeprdy1729/TIMPS-Coder-0.5B")
tokenizer = AutoTokenizer.from_pretrained("sandeeprdy1729/TIMPS-Coder-0.5B")

messages = [
    {"role": "system",  "content": "You are TIMPS-Coder v3. THINK through the root cause, FIX with complete code, VERIFY edge cases."},
    {"role": "user",    "content": "Fix: `data['user']['email']` throws KeyError when email is absent."},
]
text   = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
out    = model.generate(**inputs, max_new_tokens=700, temperature=0.1, do_sample=True)
print(tokenizer.decode(out[0], skip_special_tokens=True))

MLX (Mac Apple Silicon)

pip install mlx-lm
mlx_lm.generate \
  --model sandeeprdy1729/TIMPS-Coder-0.5B \
  --max-tokens 700 --temp 0.1 \
  --prompt '<|im_start|>system
You are TIMPS-Coder v3. THINK through the root cause, FIX with complete code, VERIFY edge cases.<|im_end|>
<|im_start|>user
Fix the race condition: two threads increment self.count += 1 simultaneously.<|im_end|>
<|im_start|>assistant
'

Training Details

Fine-tuning Configuration

Training Data

All samples formatted in ChatML with THINK → FIX → VERIFY answer structure.

Capabilities

Usage Tips

• Temperature: Keep at 0.1 — higher values increase hallucination on a 0.5B model
• Context: Include the full function/class when asking for a bug fix
• Verification: Always test generated code. Even at 88% accuracy, edge cases exist
• System prompt: Required for best results — see the Quick Start examples above

Training Code

Full training pipeline available at:
https://github.com/Sandeeprdy1729/TIMPS-Coder

License

Apache 2.0 — free to use, modify, and distribute commercially.
Base model (Qwen2.5-Coder-0.5B-Instruct) is also Apache 2.0.