Solo developer with 10+ years of
experience in Android development and the broader
developer ecosystem. Currently building IDFU on
Hugging Face — a verified-code DPO training da... Show more
namakoo
@namakoo
池澤允彦
Building IDFU — verified-code DPO training data for LLM fine-tuning
About
Language & Tools
Python, Docker, PyTorch, Transformers, TRL DPOTrainer,
Qwen2.5-Coder, HuggingFace Hub, ChromaDB, SQLite,
Stripe, Android, LoRA, Quantization
Qwen2.5-Coder, HuggingFace Hub, ChromaDB, SQLite,
Stripe, Android, LoRA, Quantization
Currently Exploring
DPO training data curation, LLM fine-tuning at 4-bit
quantization, on-device LLM agents, Python failure
pattern taxonomy
quantization, on-device LLM agents, Python failure
pattern taxonomy
Achievements
+3.46 ± 0.35 pp HumanEval improvement (3 seeds, Qwen2.5-Coder-3B) using 500 curated DPO pairs.
Built and shipped IDFU — a verified-code DPO dataset family — as a solo developer, with 8 products across 5 price tiers on Hugging Face.
Featured contributor on CoderLegion.
Built and shipped IDFU — a verified-code DPO dataset family — as a solo developer, with 8 products across 5 price tiers on Hugging Face.
Featured contributor on CoderLegion.
Fun Fact
Started in custom ROM development a decade ago, ended up curating failure patterns for code LLMs. The throughline is the same: figuring out exactly where things break.
Random Dev Quote
Most of the work in DPO training data is on the rejected side.
Contributions
Comments
- I’m all for learning as much as possible about AI.
I honestly think ...May 24, 2026
- Strong framing.
and I'd add: the judgment isn't just about evaluating ...May 17, 2026
- Wow iter a great tool.May 07, 2026
- @[Jundarer]
Fair concern — noise was the main risk going in. Three ...May 04, 2026
- The "session token issue it found on its own" line stayed with me.
Tha...May 03, 2026
chevron_left