23 stories tagged fine-tuning, newest first.
135 points by erikkannike | 95 comments
Blog post introducing a fine-tuned Qwen3-4B-Instruct model, trained on a MacBook Air, that avoids common AI writing tics and runs locally via Ollama.
609 points by jasondavies | 213 comments
Cloudflare blog post announcing Clef, a family of open-source decision models, and a new reinforcement learning fine-tuning platform.
538 points by firelex | 205 comments
GitHub repository for Jeff, a project that fine-tunes Qwen3.5 and Gemma 4 small models for fast zero-shot classification decisions.
32 points by topaztee | 8 comments
Vespper blog post launching its DOCX MCP, a fine-tuned model that lets AI agents edit Word documents via an HTML projection, claiming 3× speed and 2× lower cost.
425 points by tosh | 192 comments
GitHub repository for Kev, a tiny Jev-like family of decision models built on Qwen3.5 that users can train and run themselves.
605 points by Anon84 | 182 comments
arXiv paper showing that fine-tuning an LLM on its own sampled code outputs (simple self-distillation) improves code generation, e.g. Qwen3-30B on LiveCodeBench.
384 points by bilsbie | 93 comments
Unsloth documentation guide explaining how to fine-tune Qwen3.5 LLMs (0.8B to 122B) with LoRA, including VRAM requirements and free Colab notebooks.
334 points by Cynddl | 333 comments
arXiv paper finding that fine-tuning language models to be warm and empathetic raises error rates and makes them more sycophantic on safety-critical tasks.
376 points by trott | 126 comments
arXiv paper presenting LIMO, a model showing complex math reasoning in LLMs can be elicited with only 817 curated training samples.
452 points by felarof | 87 comments
Felafax blog post describing how they fine-tuned Llama 3.1 405B on 8 AMD MI300X GPUs using JAX, covering parameter sharding and LoRA.
376 points by majc2 | 86 comments
Blog post evaluating finetuned Mistral, Llama3 and Solar LLMs against OpenAI's GPT-4 on accuracy for structured data extraction from press releases.
437 points by aazo11 | 140 comments
GitHub repository for Dataherald, an open-sourced natural-language-to-SQL engine using LLM agents, with admin console, enterprise backend, and Slackbot services.
601 points by divan | 184 comments
Ask HN thread asking for the best current approach (as of Dec 2023) to make an LLM answer questions over a custom document set, including RAG.
324 points by dmakian | 89 comments
Substack post by David Hershey sharing tips, examples, and lessons from fine-tuning Mistral 7B to make Magic: The Gathering draft picks.
344 points by cgwu | 374 comments
Substack post by Vikram Sreekanti and Joseph E. Gonzalez arguing OpenAI's infrastructure scale makes its models too cheap for competitors to beat.
825 points by kcorbitt | 158 comments
Show HN post sharing notebooks and lessons on fine-tuning Llama 2 7B to match GPT-4 on a recipe classification task at far lower cost.
387 points by jdkee | 117 comments
A fast.ai blog post by Jeremy Howard and Jonathan Whitaker on unusual fine-tuning loss curves suggesting LLMs can memorize examples after seeing them once.
737 points by rushingcreek | 258 comments
Phind blog post announcing fine-tuned CodeLlama-34B models that score 67.6% and 69.5% pass@1 on HumanEval, beating GPT-4's reported 67%.
363 points by davidbarker | 205 comments
OpenAI blog post announcing fine-tuning support for GPT-3.5 Turbo, letting developers customize the model with their own data, plus related API updates.
372 points by 7d7n | 49 comments
Eugene Yan's long-form blog post describing seven practical patterns, such as evals, RAG, and fine-tuning, for integrating large language models into systems and products.
562 points by fzliu | 142 comments
GitHub repository for BentoML's OpenLLM, an open platform to fine-tune, serve, deploy, and monitor large language models in production.
591 points by jakedahn | 199 comments
Blog post describing how the author fine-tuned Stable Diffusion with Dreambooth to generate a painted portrait of their dog.
668 points by izzymiller | 328 comments
Izzy Miller's blog post about fine-tuning an uncensored large language model on 500,000 messages from his college friends' group chat.