{"slug": "you-fine-tuned-the-model-on-your-catalogue-and-it-only-learned-it-one-way", "title": "You Fine Tuned the Model on Your Catalogue and It Only Learned It One Way", "summary": "Serguey Asael Shinder describes how fine-tuning a model on a product catalogue taught it facts in only one direction: asked which boots are waterproof, it confidently named the wrong product even though the fact appeared hundreds of times in training. He cites the 2023 'reversal curse' research led by Lukas Berglund, which found models trained on 'A is B' sentences answer reverse questions no better than chance, and recommends writing facts in both directions, keeping facts in retrieval, and testing with questions phrased the way users actually ask them.", "body_md": "The team fine tuned a model\n\non the product catalogue.\n\nThousands of lines\n\nlike this one.\n\nThe ZR-7 is our only fully waterproof walking boot.\n\nAsk it about the ZR-7\n\nand it knew.\n\nWaterproof.\n\nWalking.\n\nBoot.\n\nThe sizes,\n\neven.\n\nThen a customer asked\n\nwhich of your boots is waterproof.\n\nIt named a different one.\n\nConfidently.\n\nThe fact was in the training data,\n\nhundreds of times.\n\nJust pointing the other way.\n\nIn 2023\n\nresearchers led by Lukas Berglund\n\ngave this a name.\n\nThe reversal curse.\n\nThey fine tuned models\n\non invented sentences\n\nof the form A is B,\n\na made up person\n\nis the director of a made up film.\n\nAsked who directed the film,\n\nthe models did no better than chance.\n\nThey found it in the big models too,\n\nwith real people.\n\nGPT-4 could say who Tom Cruise's mother is\n\nabout four times in five.\n\nGiven her name\n\nand asked who her son is,\n\nabout one time in three.\n\nTraining on a sentence\n\nteaches the model to continue it\n\nin the order it was written.\n\nIt does not, by itself,\n\nstore a fact you can enter\n\nfrom either side.\n\nPeople do that without noticing.\n\nModels, mostly, do not.\n\nThe odd part\n\nis that inside the prompt\n\nit is fine.\n\nPut the sentence in the context,\n\nask the reverse question,\n\nand the model answers it,\n\nbecause now it is reading,\n\nnot remembering.\n\nSo think about the direction\n\nof every question your users will ask.\n\nCustomers rarely start with the code.\n\nThey start with the need,\n\nthe size,\n\nthe symptom.\n\nIf you fine tune,\n\nwrite the facts both ways,\n\nand in the ways people actually ask.\n\nBetter,\n\nkeep the facts in retrieval,\n\nwhere the model reads them\n\nat the moment it answers,\n\nand use fine tuning\n\nfor tone and format.\n\nAnd test in the direction\n\nyour users come from.\n\nIf every test question\n\nstarts with the product name,\n\nyou have only checked\n\nthe half of the fact\n\nyou already knew it had.\n\n– Serguey Asael Shinder", "url": "https://wpnews.pro/news/you-fine-tuned-the-model-on-your-catalogue-and-it-only-learned-it-one-way", "canonical_source": "https://dev.to/sergueyasaelshinder/you-fine-tuned-the-model-on-your-catalogue-and-it-only-learned-it-one-way-3fek", "published_at": "2026-11-02 07:43:20+00:00", "updated_at": "2026-11-02 07:52:33.279710+00:00", "lang": "en", "topics": ["large-language-models", "machine-learning", "ai-research", "ai-tools"], "entities": ["Serguey Asael Shinder", "Lukas Berglund", "GPT-4", "Tom Cruise"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/you-fine-tuned-the-model-on-your-catalogue-and-it-only-learned-it-one-way", "markdown": "https://wpnews.pro/news/you-fine-tuned-the-model-on-your-catalogue-and-it-only-learned-it-one-way.md", "text": "https://wpnews.pro/news/you-fine-tuned-the-model-on-your-catalogue-and-it-only-learned-it-one-way.txt", "jsonld": "https://wpnews.pro/news/you-fine-tuned-the-model-on-your-catalogue-and-it-only-learned-it-one-way.jsonld"}}