IndustryLLM Uses Failure-Driven Training for Industrial Procurement

Summary

Industrial procurement must connect informal buyer language and sparse marketplace attributes with authoritative engineering standards while maintaining strict safety tolerances. The paper introduces IndustryLLM, an open-weight industrial language model initialized from Qwen3.5-35B-A3B-Base, with 35 billion total parameters and about 3 billion activated per token; its vision encoder is frozen. Its adaptation recipe combines continued pre-training and supervised fine-tuning rather than relying only on generic text scaling. The training corpus is described as roughly 100 billion tokens, including 5 billion tokens of national standards and technical archives, 10 billion tokens of de-identified industrial transaction and inquiry records, and 60 billion tokens of general replay. The authors also reconstruct an estimated 20 billion-token domain subset through multi-register rewriting across 10 genres and 8 styles, confidence-routed factual editing, and error-targeted question-answer synthesis. Examples include correcting colloquial material names, expanding ambiguous standards codes, and resolving conflicting dimensional specifications. For deployment, IndustryLLM uses an evidence-gated constraint-evaluation interface with three-valued logic, so unsupported product evidence remains unknown rather than being treated as satisfying a constraint. Offline procurement-query structuring improved by 2.97 percentage points in exact match in No-Think mode, with a 95% confidence interval of 2.11 to 3.86 points. In randomized online production A/B tests, the paper reports a 4.25% GMV increase, an 8.3% increase in satisfied inquiries, and latency falling from 6-7 seconds to 1.5 seconds. The model weights and configurations are released on Hugging Face.