AI repos for mL engineers and data scientists
Open-source AI projects worth it for people who train, fine-tune and evaluate models and study how they work, each with what it does for you. 33 repos, most relevant first.
-
transformers
Standard API to run, fine-tune and evaluate pretrained text, vision, audio and multimodal models from the Hub.
-
LLMs-from-scratch
Build tokenizer, attention, GPT pretraining and finetuning in PyTorch to understand LLM internals.
-
autoresearch
Compact GPT training script with Muon and AdamW, plus an overnight agent loop that cuts validation loss.
-
minimind
Run pretraining, SFT, LoRA, DPO and GRPO on a 64M model that fits a single GPU.
-
Kronos
Fine-tune and evaluate pre-trained candlestick forecasting transformers on your own market data.
-
timesfm
Pretrained time-series foundation model for zero-shot forecasting with quantiles and covariates, useful to study and benchmark.
-
cosmos
Use open world-model checkpoints, notebooks and benchmarks for video generation, robot policies and synthetic driving data.
-
reasoning-from-scratch
Implement GRPO, inference-time scaling and distillation in PyTorch on a Qwen3 base model, following a book.
-
airllm
Run and compare 70B+ open models on a 4GB GPU by loading one layer at a time.
-
VoxCPM
Open 2B text-to-speech model with voice design and cloning to study or fine-tune.
-
modular
MAX serves models via an OpenAI-compatible endpoint, and its kernel code shows how GPU operations are written.
-
Unlimited-OCR
Run and evaluate an open document vision-language model with transformers, vLLM or SGLang.
-
lingbot-map
Streaming feed-forward 3D reconstruction model with checkpoints and benchmarks to study or build on.
-
GLM-5
Open-weights coding model with serving guides and ms-swift fine-tuning recipes for SFT, PPO and GRPO.
-
GLiNER
Train, fine-tune and run small zero-shot NER models, including bi-encoder and relation extraction variants.
-
tribev2
Pretrained multimodal Transformer predicting fMRI responses, with training code for brain-encoding research.
-
HyperQwen
Apply vLLM patches and model preparation to serve a 27B Qwen fast on one 24 GB GPU.
-
OSWorld-V2
Benchmark and inspect trajectories of GUI agents against published results.
-
glance-speedlab
Benchmark PyTorch MPS against MLX 8-bit for Qwen3-VL-2B and reproduce preregistered latency experiments.
-
ComfyUI
Fine-grained control over diffusion models and parameters, with an API to plug workflows into software.
-
whisper
Study and benchmark a widely used multilingual speech recognition model across sizes.
-
RuView
Study WiFi channel state information for presence, breathing and heart-rate estimation on cheap ESP32 hardware.
-
ai-engineering-from-scratch
Implement backprop, tokenizers and attention by hand to deepen understanding of model internals.
-
AI-For-Beginners
Compare the same neural network in PyTorch, Keras and TensorFlow notebooks to learn fundamentals.
-
dive-into-llms
Work through fine-tuning and a mini R1 distillation of a small math-reasoning model.
-
gallery
Compare open models on your own device with benchmarks and sampling settings.
-
needle
Small 8-29 MB model for tool calling and embeddings, useful for edge deployment study.
-
ACE-Step-1.5
Open music generation model running across Mac, AMD, Intel and CUDA to study or fine-tune.
-
supertonic
Compact 99M-parameter TTS model running via ONNX Runtime without a GPU.
-
PixelRAG
Explore visual document retrieval with a fine-tuned vision-language embedding model and a hosted Wikipedia index.
-
microduck
See how sim-trained reinforcement learning policies are exported to ONNX and run in a 50 Hz loop on real hardware.
-
StartLux-Decision
Open 0.8B-27B decision models with calibrated outputs to evaluate or self-host.
-
AutomationBench
Use partial-credit scores as a reward signal when tuning agents on business tasks.