$ articles --topic=fine-tuning
Fine-Tuning
16 AI articles about Fine-Tuning.
Accelerating Reinforcement Learning Weight Refit with the Reshard CollectiveLearn how the Reshard collective speeds up weight refit in reinforcement learning post-training, where updated policy weights must be synced from training GPUs to separate inference GPUs in a disaggregated setup. The article covers why this transfer becomes a bottleneck and how the new approach reduces it.Inference · via NVIDIA Technical Blog · 1d ago1 minZero Mismatch, Simple Async RL: REINFORCE Zero, DeepSeek-V4.1 and Kimi K3.1Explains how asynchronous reinforcement learning can train on stale model weights, and how aligning rollout and training kernels bit for bit, together with unclipped importance sampling, corrects that lag. Covers zero-mismatch RL applied to DeepSeek-V4.1-Flash, Kimi K3.1 and AMD MI355X hardware.Open Models · via vLLM · 1d ago1 minPalette: teaching a small model to build decks in your design languageOpen Models · via Hugging Face · 4d ago1 minCustom Voices With Qwen3-TTSLearn how to build custom TTS voices with Qwen3-TTS by blending voice embeddings from a designed voice and a real recording, and how to add emotional instructions like sad or whispering to a cloned voice by pairing a Base speaker embedding with the CustomVoice model.Open Models · via Armin Ronacher · 9d ago9 minBuilding Reasoning Models from the Ground Up: A Full Video CourseThis course walks through building reasoning models from scratch in Python and PyTorch. It covers how to evaluate such models, scale compute at inference time, train them with reinforcement learning, and apply distillation. The full video series runs over 11 hours of coding lessons.Evals · via Sebastian Raschka · 14d ago1 minIntroducing our Training APISpindle is an open-source API for training and fine-tuning open models. You can deploy a server with a single command and run any Tinker-compatible training code against it.Open Models · via Modal · 16d ago1 minTorchTitan Update: New SOTA Models, Performance at Scale, TitanRLThis update covers TorchTitan's new open-weights model support, including Qwen3.5, Kimi, and DeepSeek-V4, and its TitanRL reinforcement learning stack that reuses one model across training and vLLM inference. It also reviews MoE performance gains, MLPerf and long-context results, ecosystem integrations, and the FlexShard data parallelism library.Open Models · via PyTorch · 17d ago7 minGet started with LLM training on ModalLearn how to begin training large language models with reinforcement learning on Modal. The guide introduces Modal Dojo, a library that provides building blocks, tuned defaults, and observability tools for production-grade LLM training.Inference · via Modal · 17d ago1 minHow To Use Decision Models for EvalsDecision models offer fast, cheap classification for automated evals, but the default 0.50 threshold often leaves performance on the table. This post compares several models on cost and accuracy, shows how to tune thresholds for F1 with precision and recall minimums, and explains how calibration makes scores more useful.Evals · via Hamel Husain · 18d ago8 minTogether Post-Training: frontier-grade model shaping, the way that fits youExplains how to post-train open LLMs such as Qwen, GLM and Kimi into custom models you own, using supervised fine-tuning or reinforcement learning. You write the training loop while Together handles the underlying infrastructure.Open Models · via Together AI · 19d ago1 minGoogle DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detectionLearn how Google DeepMind and GovTech Singapore built a fine-tuned scam detection model using lightweight Gemma models that run on-device. The approach intercepts conversational scams in private chats without sending any data to the cloud.Open Models · via Google DeepMind Blog · 24d ago1 minFocusing on Post-TrainingThe author explains why post-training existing open-weight large language models may be a worthwhile investment. Fireworks' Ember-1 serves as an example of a model that reasons with greater token efficiency.Open Models · via Sebastian Raschka · 1mo ago1 min