sft-fab-3b

XCombinator — sft-fab-3b

⚠️ Post-deadline upload notice. This Hugging Face repository was published after the Zero One Hack_01 submission deadline (2026-05-31 10:00 CET), solely to give judges download access. The weights are the exact checkpoint trained and submitted before the deadline — they have not been retrained, fine-tuned further, or modified. Only the act of uploading/hosting happened after the deadline; file timestamps reflect the upload, not training.

Full fine-tune of Qwen/Qwen2.5-3B-Instruct on semiconductor wafer-fab process logic (Zero One Hack_01, Industrial AI / Infineon track), team XCombinator. Model-size point — 3B base, full data (~3 epochs, 4-GPU FSDP).

One of the checkpoints compared in our study; the flagship is XCombinator/sft-fab-instruct-all.

Prompt format

Unified JSON format: a system prompt (task + output schema) + a numbered user sequence → one JSON answer ({"reasoning": "...", "steps": [...]} for next-step/completion; {"reasoning": "...", "valid": bool, "rule": "RULE_..."|null} for anomaly). Build the exact messages with zo_train.prompts.build_messages from the project repo, then apply the tokenizer chat template. See the flagship model card for a full from_pretrained snippet.

Evaluation (MOSFET labeled eval, n≈200)

Full study + all checkpoints: the project repo and submissions/XCombinator/REPORT.md.

Notes

• Full fine-tune (not a LoRA adapter) — loads directly with AutoModelForCausalLM.from_pretrained.
• Trained on Leonardo (CINECA) A100 via a deterministic data factory over the organizer grammar.