模型後訓練與企業 AI
9B Model Learns Catalog Review With $500 of GRPO Training, Outperforms Frontier APIs on a Single Workflow
Fermi Sense trained a 9B open model to achieve 87.3% of the attainable score in a gradable e-commerce catalog environment using two GPUs and about $500. The results suggest that enterprise rules can be encoded into small models through reinforcement learning, but the evaluation data, model weights, and full reproducibility package have not yet been released.
Fermi Sense has published a vertical-model experiment in which the team rebuilt e-commerce product catalog review as a repeatable environment. Using images, descriptions, a category tree, and tool outputs, the model must determine whether a product’s category and attributes are correct. Each decision is scored according to fixed rules, after which GRPO is used to generate rollouts and update a 9B open model—instead of continually stuffing the full set of enterprise rules into prompts for frontier models.
The team used two RTX PRO 6000 GPUs, with one generating trajectories and the other performing gradient updates. Using the open-source prime-rl framework, it ran 1,000 optimization steps over roughly three and a half days at an estimated GPU cost of about $500. The final model achieved a raw score of 0.626 on the strict evaluation, equivalent to 87.3% of the attainable ceiling defined by the authors. The untrained base model reached 64.2%, while the best frontier API configuration achieved 76.9%. After roughly 250 training steps, the model had already surpassed the frontier-model range. The authors estimate inference costs at $0.50 per 1,000 products—about 40 times less than the cheapest comparison API—while the best frontier configuration cost $34 per 1,000 products.
The technical takeaway is not that the 9B model has stronger general-purpose reasoning, but that stable classification rules, tool-use patterns, and exception handling can be compressed into its weights, eliminating the need to resend lengthy prompts with every request. This approach is well suited to high-frequency workflows whose outputs can be mechanically verified and whose rules remain relatively stable. If work quality can be assessed only through subjective judgment, however, GRPO’s reward function may instead entrench flawed proxy metrics.
For now, the figures are primarily vendor-reported. The test set, post-training weights, random seeds, and complete cost breakdown have not been released, and the experiment covers only a single catalog workflow. Engineering teams should next examine data drift, retraining costs after rule changes, reward hacking, and the small model’s ability to generalize across rare categories and different stores’ taxonomy systems.