PostTrainBench releases 58.4GB of agent fine-tuning traces, 151k downloads in a month
maksym_andr · x · 2026-11-13
The AI Safety group at ELLIS Institute Tübingen and MPI-IS released PostTrainBench-Trajectories on Hugging Face: 58.4GB of agent traces, already at 151k downloads last month.
- Traces come from PostTrainBench, where a CLI agent gets a base LLM, an eval script, and 10 hours on one H100 80GB to autonomously improve the model on a target benchmark via any post-training strategy (SFT, LoRA, RLHF, data-generation prompting, etc.).
- Benchmarks covered include AIME 2025, GPQA, HumanEval, GSM8K, HealthBench, and BFCL; base models include Qwen3-1.7B/4B-Base, SmolLM3-3B, and Gemma-3-4B-PT.
- Released under Apache-2.0, useful for distilling agent trajectories and studying learned post-training strategies.
More from coding & agent
- Solo founder uses Claude Code to run SEO service, new client clicks up 78% in 2 weeks — doooyle · 2026-11-13
- Decision models act poorly but teach well: GRPO teacher lifts Qwen 1.5B to 82.9% on ALFWorld — omarsar0 · 2026-11-13
- Google Antigravity's subagent commands compared: /swarm, /boost, /spawn and when to use each — fhinkel · 2026-11-13
- Math, Inc. ships executable Lean model of Linux WireGuard handshake — andriy_mulyar · 2026-11-13
- One-shot prompts build a Notion-style markdown editor and JSON viewer with ui sdk — shadcn · 2026-11-13
- Agent coding burns 1-2B tokens a day: Reddit dev does the math on a $1,000+/month API bill — JamesGooning · 2026-11-13