{"slug": "datasets-and-the-right-models", "title": "Datasets and the right models", "summary": "A developer is seeking advice on which model excels at selection-with-abstention tasks, where the model must either pick a candidate with a confidence tier or abstain from outputting anything. The user has tried Gwen and Mistral but found them unsuitable for this specific structured output task, which involves fine-tuning a small instruct model via LoRA on a small corpus.", "body_md": "Hey I was hoping for advice on which model can handle the sort of data sets I want to FT on. I’d like to give my model a corpus of information and train them to say nothing when they lack confidence. I’ve been using Gwen for this, and tried our Mistral to leave it takes date of another kind. I can’t even train it on what I need it to know. So which model excels at It’s a **selection-with-abstention** tasks, not generation. Each row gives the model a *cue* (a conversation turn) plus a **numbered candidate slate** of informative chunks, and the target output is a short structured string that either **picks an index with a confidence tier** or **abstains**:\n\n│ -`— surface candidate 3, confident │ -`\n\n— surface it, hedged │\n\n│ - — abstain, surface nothing │\n\n│ │\n\n│ Format is completion-style SFT on a small instruct model (LoRA), small corpus (a few hundred real items, augmented).", "url": "https://wpnews.pro/news/datasets-and-the-right-models", "canonical_source": "https://discuss.huggingface.co/t/datasets-and-the-right-models/176969#post_1", "published_at": "2026-06-18 23:44:01+00:00", "updated_at": "2026-06-19 00:09:23.322935+00:00", "lang": "en", "topics": ["large-language-models", "machine-learning", "ai-research"], "entities": ["Gwen", "Mistral", "LoRA"], "alternates": {"html": "https://wpnews.pro/news/datasets-and-the-right-models", "markdown": "https://wpnews.pro/news/datasets-and-the-right-models.md", "text": "https://wpnews.pro/news/datasets-and-the-right-models.txt", "jsonld": "https://wpnews.pro/news/datasets-and-the-right-models.jsonld"}}