StepFun open-sources onPanda: token-level LLM annotation and inspection tool
teortaxesTex · x · 2026-09-23
StepFun open-sourced onPanda, the tool it uses internally for LLM data annotation and model inspection — born from diyerxx's two-year effort building a token-level visualization & control web app (mobile-friendly). Workflow: find an error, correct the token, let the model continue.
Data annotation
- 52% lower median annotation time vs. manual post-editing
- SFT + preference data in one workflow, high on-policy fidelity (ΔPPL <1% vs. the model's resampling baseline)
- Precise token-level supervision with paired positive/negative examples; agent-trajectory annotation across image, audio, and video
Model inspection & debugging
- Inspect token probabilities and top-k alternatives, steer decoding token by token
- Explore SVG generation, web development, and agent tasks directly in the browser
Online demo and paper linked in the post.
Related event: StepFun Open-Sources onPanda for Faster Alignment Data Labeling(3 posts)→
More from coding & agent
- Browser Use Bench v2: GPT-6 Sol scores 66.9, beating Opus 5.5 at 3.5x lower cost — airesearch12 · 2026-09-23
- China's AI coding assistant disables features after code data loss found in security audit — pstAsiatech · 2026-09-23
- Rogo Co-founder: Finance Is the Best Fit for 10,000-Agent Swarms Where One Insight Is Worth $50k — rohanpaul_ai · 2026-09-23
- Miles Brundage notes Claude Code mode lets you 'show more' but not 'show less' — Miles_Brundage · 2026-09-23
- jev-gc: reversible context garbage collection for long-running AI agents — Maleficent_College57 · 2026-09-23
- Devs call Effect + Alchemy + Cloudflare a 'superpowers stack' made 100x easier by AI coding — samgoodwin89 · 2026-09-23