LLMs & ML

google-research/kauldron

Google's JAX training framework built for fast research iteration — Kauldron is a modular JAX/Flax training library from Google Research that lets you mix any dataset loader, model, or optimizer while customizing every inner layer. It emphasizes a Colab-first workflow with built-in plotting, profiling, and post-mortem debugging to speed up research iteration.

Modular, scalable library to train ML models

FrigadeHQ/hob

Local AI desktop app that auto-picks the right open-weight model for your Mac — Hob is a free, open-source desktop app that runs open-weight models like Gemma and gpt-oss locally via llama.cpp, offering offline, private alternative to Claude and ChatGPT. It automatically selects the best frontier model that fits your hardware and sets up everything, letting you fill PDFs, analyze documents, and chat without sending data to the cloud.

An easier way to use AI on your own computer. A free, open-source alternative to Claude and ChatGPT that runs open-weight models like Gemma and gpt-oss locally on Mac, Windows and Linux. Private, and works offline.

nu-dialogue/j-moshi

J-Moshi brings full-duplex real-time dialogue to Japanese speech AI — J-Moshi adapts Kyutai's 7B full-duplex speech dialogue model Moshi to Japanese by fine-tuning on Japanese conversational speech data, enabling natural overlapping turn-taking and backchannels in real time. The repo ships two pretrained variants, one enhanced with synthetic multi-stream TTS data, plus demo samples and training code for researchers in spoken dialogue systems.

J-Moshi: A Japanese Full-duplex Spoken Dialogue System

menloresearch/cyclotron

Open training pipeline for Asimov 1 humanoid locomotion — Cyclotron is Menlo Research's Isaac Lab extension for training and deploying locomotion policies on the open-source Asimov 1 humanoid robot, using PPO and adversarial motion priors. It covers the full workflow from training through export, viewing and benchmarking with simple commands, aimed at anyone building or customizing policies for sim-to-real humanoid control.

Official reinforcement learning training framework for Asimov humanoid robots, built on Isaac Lab.

thomasunise/casecite

Self-hosted legal RAG platform that keeps documents on your own infrastructure — CaseCite bundles nine tools — document chat, deep research against your corpus and CourtListener, contract review, and tabular review across hundreds of contracts — behind one login. It's self-hosted and BYOK: documents, search index, and chat history stay on your servers, with no telemetry to the maintainers. Aimed at firms wanting legal RAG without sending data to a third-party SaaS.

CaseCite is an open-source, self-hosted legal RAG caselaw research & document analysis platform.

ryota-komatsu/slp2025

Conference materials and demo code for an audio language model survey — This repo bundles slides and runnable demo scripts from academic talks on audio/speech language models, including a hands-on walkthrough connecting Whisper encoder and Llama 3.2 via a lightweight adapter for instruction following. It suits researchers or students wanting reproducible speech-LLM pretraining and instruction-tuning code alongside the presented slides.

Survey of audio language models

Arnie936/llm-prompting-tests

A German-language prompt suite that stress-tests LLMs on real system logic — This repo collects 22 demanding German-language prompts for testing LLMs, coding agents, and agentic models. Rather than simple Q&A, prompts like the beehive simulation task require correct module planning, decoupled tick/render loops, and resource-pool logic in a single-file HTML app—useful for anyone evaluating whether a model can reason about coupled systems, not just produce surface-level output.

Prompting tests for evaluating LLMs, coding agents, and agentic models

Atotti/ja-speech-llm

A Full Training Pipeline for a Japanese Speech-Understanding LLM — This repo provides an end-to-end pipeline that connects a frozen Whisper-large-v3 encoder to an LLM-jp 8B decoder via a trainable adapter, enabling Japanese speech chat and ASR. It covers ASR pretraining, multi-task SFT, DPO alignment, and a Gradio demo, with released checkpoints and multiple training modes (adapter-only, LoRA, full).

FT-LLM2026で開発した日本語SpeechLLMの学習コード

scienthoon/jev-ood-calibration

Testing Whether Jev's Confidence Scores Stay Honest Out-of-Distribution — An independent, reproducible (~$0.06) calibration audit of TypeSafe's Jev model, comparing its probability estimates on public benchmarks it likely trained on versus 900 novel rule-based tickets it cannot have seen. Finds Jev stays well-calibrated in-distribution but grows confidently wrong on unseen rules, with the miscalibration's sign flipping by question type — useful for anyone evaluating Jev's 'epistemically honest probabilities' claim before routing decisions to it.

Independent calibration test of TypeSafe's Jev on a task it cannot have seen: 900 rule-generated support tickets (choice / score / boolean) plus 3 public benchmarks via Vercel AI Gateway. Raw responses, ECE with noise floor, temperature refit, per-type sign of miscalibration. Reproducible for ~$0.06.

AnthusAI/Jev-Calibration

Fixing Jev's Overconfidence: Isotonic Regression Beats Platt Scaling — This repo tests whether Jev's stated confidence matches real accuracy on 8,801 sentiment examples. Raw probabilities had an expected calibration error of 0.117; Platt scaling cut it to 0.052, but isotonic regression dropped it to 0.008 and worked with as few as 20-100 calibration examples. It's a practical follow-up for anyone building LLM-based classifiers who need trustworthy confidence scores, not just correct answers.

Does Jev's confidence mean what it says? Calibrating Jev (TypeSafe System One) with Platt scaling and isotonic regression.

furuya02/sagemaker-embedding-ft-visual-inspection

Fine-tuning loses to better lighting in zero-defect visual inspection — An experiment measuring anomaly detection for visual inspection trained on zero defect images, using a 67-photo rubber duck dataset (CC BY 4.0) bundled in the repo. It finds fine-tuning actually worsens the usable threshold range, while controlling lighting and switching to DINOv2 embeddings gives far bigger gains, even though AUROC looks identical at 1.000 in all cases.

不良画像を1枚も使わない外観検査で、ファインチューニングは割に合うのか実測した記録。アヒル67枚のデータセット(CC BY 4.0)を同梱

kmizu/syntax-locality-experiments

Does verbose syntax help LLMs generate code more accurately? — Paired experiments testing whether redundant terminators (like `end func n_abcdefgh` vs `}`) improve LLM accuracy when generating code in unfamiliar miniature grammars. Built in pure Scala/JDK with no external SDKs, it produces reproducible HTML reports—useful for researchers studying context locality and prompt-driven code generation.

Paired experiments on supplied miniature grammars: reading, generation, and full-program parsing. Reproducible HTML reports.

cm-nakamura-shogo/blog-deepseek-v4p1-1017-self-hosting

Self-hosting DeepSeek V4.1 Flash on AWS G7e with SGLang and DSpark — A reproducible verification log for running DeepSeek-V4.1-Flash-1017 on AWS G7e instances via SGLang/ECS/Terraform. It documents DSpark speculative decoding boosting throughput roughly 2x (up to ~55 tok/s), an Engram offload comparison, a failed vLLM attempt, and capacity shortages across regions—useful for anyone self-hosting large MoE models on new AWS GPU instances.

DeepSeek V4.1 Flash 1017 on AWS G7e with SGLang (ECS, Terraform) - blog verification code