RL fine-tuning Qwen3.5-0.8B for under $20 on HF Jobs lifts pass@1 from 10% to 36%
LysandreJik · x · 2026-10-20
- A practitioner used RL (AsyncGRPO) via HF Jobs to fine-tune Qwen3.5-0.8B on a target agent harness, raising pass@1 from 10% to 36% on a 50-task held-out eval set (peaking at 0.36 around step 50, dipping to 0.32 at step 100).
- Mean tool calls per task dropped from 14.6 to 8.2, nearly halving agent steps.
- Everything fits on a single 24GB GPU, costs under $20, and requires no synthetic trajectory generation. Shared/endorsed by HF's Lysandre.
More from coding & agent
- Color Scheme Generator MCP Server offers 8 palette methods for AI design workflows — modelcontextprotocol · 2026-10-20
- Klaviyo MCP Pack wraps email marketing API for AI assistants — modelcontextprotocol · 2026-10-20
- Ethan Mollick: new Claude Projects shines as a coordinator-plus-agents interface for multi-agent work — emollick · 2026-10-20
- Yacine: After AI I commit less code but write far more words — yacineMTB · 2026-10-20
- Google engineer releases a Skill that turns any code diff into an interactive visual explainer for PR review — ykdojo · 2026-10-20
- From hand-coding to "approve once": human sign-off is the new bottleneck — ThePeterMick · 2026-10-20