Projects Recently Starred by Experts

Starred 14 hours ago by

Didier Lopes(Founder of OpenBB). Also starred by

Tomas Valenta(Cofounder of E2B),

Travis Fischer(Founder of Agentic),

Jinze Bai(Research Scientist at Alibaba Qwen), and

36 more.

gym by openai

0.0%

37k

RL toolkit for reinforcement learning algorithm development

Standard API for learning algorithm/environment communication.
Includes a standard set of environments.
API models environments as simple Python env classes.

Created 10 years ago

Updated 3 months ago

Starred 18 hours ago by

David Cournapeau(Author of scikit-learn). Also starred by

Dominik Moritz(Research Scientist at Apple; Professor at CMU).

flint-chart by microsoft

163.6%

489

AI-powered visualization language for creating expressive charts

Compiles simple, human-editable specs into optimized Vega-Lite, ECharts, or Chart.js visualizations.
Leverages 70+ semantic types and automatic layout for data-driven chart generation.
Integrates with AI agents via an MCP server for interactive chart creation and validation.
Provides a JS/TS library and Python preview for seamless integration into applications.

Created 1 month ago

Updated 1 day ago

Starred 19 hours ago by

Pawel Garbacki(Cofounder of Fireworks AI). Also starred by

Chaoyu Yang(Founder of Bento),

Jeff Hammerbacher(Cofounder of Cloudera),

Luis Capelo(Cofounder of Lightning AI), and

2 more.

tokenspeed by lightseekorg

2.3%

Speed-of-light LLM inference engine

High-performance inference engine optimized for production agentic workloads.
Combines TensorRT-LLM performance with vLLM usability via static compilation and layered kernels.
Features advanced scheduler, pluggable kernels, and fast Multi-head Latent Attention (MLA) on Blackwell.
Showcases novel runtime design with compile-time safe KV cache management for LLM agents.

Created 2 months ago

Updated 14 hours ago

Starred 20 hours ago by

Didier Lopes(Founder of OpenBB).

natural_language_autoencoders by kitft

2.5%

893

Explain LLM activations with natural language autoencoders

Maps activation vectors to text and reconstructs vectors from natural language descriptions.
Uses fine-tuned LMs (AV/AR) for unsupervised analysis and generating human-readable explanations.
Built on Miles for RL training and SGLang for efficient inference serving of activation embeddings.
Provides pre-trained checkpoints for Qwen, Gemma, and Llama models for immediate use.

Created 2 months ago

Updated 1 month ago

Starred 21 hours ago by

Anton Bukov(Cofounder of 1inch Network).

jacobian-lens by anthropics

N/A

811

Interpretability tool for understanding language model activations

Decodes internal model activations into probable output tokens using Jacobian-based linear transport.
Analyzes transformer decoder models, supporting HuggingFace architectures and text modality.
Implements a reference Jacobian lens for research, enabling visualization of internal model reasoning.
Predicts salient vocabulary (e.g., "nose") from activations even when absent in the input prompt.

Created 6 days ago

Updated 6 days ago

Starred 1 day ago by

Wing Lian(Founder of Axolotl AI).

awesome-on-policy-distillation by chrisliu298

8.4%

491

On-policy distillation techniques for LLM training and alignment

Trains student LLMs on self-generated trajectories with teacher supervision, reducing distribution gaps.
Supports self-distillation, black-box methods, and cross-tokenizer alignment for flexible model training.
Integrates with frameworks like TRL and KDFlow, offering optimizations for efficiency and multimodal tasks.
Widely adopted in industry and research for LLMs, agents, and cross-modal applications.

Created 3 months ago

Updated 1 day ago

Starred 1 day ago by

Georgios Konstantopoulos(CTO, General Partner at Paradigm). Also starred by

Phil Wang(Prolific Research Paper Implementer).

Self-Distillation by idanshen

1.1%

645

Continual learning for foundation models via self-distillation

Acquire new skills and knowledge without performance degradation, outperforming supervised fine-tuning.
Uses on-policy reinforcement learning with a demonstration-conditioned model acting as its own teacher.
TRL-based code for reproducing research on self-distillation and mitigating catastrophic forgetting.
Enables sequential skill accumulation from demonstrations without performance regression.

Created 5 months ago

Updated 3 months ago

Starred 1 day ago by