📈DeepScaler
Democratizing Reinforcement Learning for LLMs
⭐ Popular: 6k stars, gaining about 18 a week
View on GitHub ↗repo profile
momentum
durability
bus factor = how many people it takes to cover more than half the commits (6 months). 1 is a solo project; higher means the work is spread across a team. top-author share is the single busiest author's slice of those commits.
since we covered it
why it's a big deal
- Provides an open framework for training language agents with reinforcement learning, so researchers can move the same agent code from evaluation into training without rewriting it.
- Ships 60+ integrated benchmarks across math, coding, QA, search, vision-language, and agentic tasks, giving teams a common baseline to test agents against.
- Lets the same setup drive real CLI harnesses like Claude Code, Codex, and opencode, so work is not locked to a single agent framework.
under the hood
- Written mostly in Python (72 percent of the codebase) with Jupyter Notebooks (26 percent), following a pipeline of run agent, collect traces, compute rewards, update model.
- Splits into a Workflow Engine for parallel rollouts, a Model Gateway that captures token IDs and logprobs, a transform pipeline for trajectory grouping, and a pluggable training backend.
- Supports GRPO, REINFORCE, RLOO, SFT, and on-policy distillation, with training backends (verl, tinker, fireworks) and sandboxes (Docker, Daytona, Modal, local) switchable by flag.
Radar summary, generated from the project's public sources
star history
- PR#2 2k 2025-02-19
- now 6k + 4k since first covered
curve is sampled from GitHub's star history, plus our own daily readings since we covered it; the dashed stretch is before we first covered it, the solid line since. figures at coverage are the numbers we printed then (approx.), current count is live.
understory
Quietly building: more output than attention, for now.
- output, commits & releases
- clout, star velocity
output = commits & releases; clout = star velocity, both 0 to 100 monthly indices; the gap where output runs above clout is the understory. The understory →
covered in
-
Democratizing Reinforcement Learning for LLMs
similar projects
compare these →- ☁️ skypilot
1.8× the stars
The AI Compute Platform for frontier teams. SkyPilot turns fragmented AI compute into one AI supercomputer, so frontier AI teams build custom intelligence faster.
10k ACTIVE - 📖 ai-agent-book
6.6× the stars
AI Agent PDF
38k ACTIVE - 🦌 DeerFlow
14× the stars
An open-source long-horizon SuperAgent harness that researches, codes, and creates. With the help of sandboxes, memories, tools, skill, subagents and message gateway, it handles different levels of tasks that could take minutes to hours.
80k ACTIVE
comments
Sign in with GitHub to add your blip on DeepScaler.