HuggingChat's ML Intern trains a LoRA from one prompt in ~75 minutes
Gradio · x · 2026-09-25
Gradio demos HuggingChat's new "ML Intern": from a single prompt it renders a dataset, runs base-model baselines, trains with ai-toolkit on one A100 (2,000 steps, 2.2 s/step, 75 minutes), picks checkpoints on a fixed subset, evaluates against ground truth, and writes the model card. Users can upload a transparent PNG and get a new view in 25 seconds, with a toggle to rerun the same prompt and seed without the LoRA to compare.
More from coding & agent
- Microsoft Foundry Agent Service adds native voice agents in public preview — lee_stott · 2026-09-25
- Dev Proposes "Evidentiality Framework": Provenance Markers for Every LLM Claim — jzesbaugh · 2026-09-25
- New Paper Quantifies When Rubric-Based RL Genuinely Improves Models vs. Hacks the Verifier — heghbalz · 2026-09-25
- Fighting agent scope drift: team adds Jev monitor to Hermes-Agent to keep agents on task — alexcovo_eth · 2026-09-25
- Connectome: open-source agent host that folds AI memories in layers like humans do — repligate · 2026-09-25
- Claude Code builds a full 3D survival game in one day: 2km island, 44k code-generated trees — majidmanzarpour · 2026-09-25