# Jiayi-Pan/TinyZero

Minimal reproduction of DeepSeek R1-Zero

Repository: https://github.com/Jiayi-Pan/TinyZero
Canonical: https://ross.abutalabs.com/products/tinyzero
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-02-27T05:42:07+00:00

## Health v2 (maintenance only)
Score: 52/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 69, release rhythm 35, longevity 42
- inputs: {"age_days": 589, "days_push": 187, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 13223, forks 1577 (observed 2026-08-28T04:11:02.613023+00:00)

## What it is
TinyZero is a minimal reproduction of DeepSeek R1-Zero, showing that a 3B base language model can develop self-verification and search abilities through reinforcement learning on countdown and multiplication tasks. It is built on top of the veRL library and is now deprecated in favor of veRL itself.

## Use cases
- reproduce deepseek r1-zero training on a small budget
- train a small language model with reinforcement learning to learn reasoning
- run RL experiments on countdown and multiplication tasks
- observe emergent self-verification in a 3B base model
- learn how to use veRL for LLM RL training
- experiment with RL post-training of Qwen models

## When to choose
- you want a cheap, minimal reference implementation of R1-Zero-style RL training
- you are studying how reasoning abilities emerge via RL in small base models
- you want a starting point built on veRL for reasoning RL experiments

## When to avoid
- you need an actively maintained RL training framework - use veRL directly instead
- you want production LLM training infrastructure
- you cannot access GPUs - training requires multi-GPU setups for 3B+ models

## Facets
- artifact type: library
- maturity: abandoned
- function: llm-training, reinforcement-learning, machine-learning
- domain: large-language-models, deep-learning, reinforcement-learning, artificial-intelligence
- platform: python
- tags: deepseek-r1-zero, verl, reasoning, rl-training, research-reproduction, countdown-task, vllm, gpu, linux

## Member repositories
- Jiayi-Pan/TinyZero (main) score 52

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:11:02.613023+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:13:25.776209+00:00, confidence not recorded.
  - readme: https://github.com/Jiayi-Pan/TinyZero (fetched 2026-08-28T04:11:02.613023+00:00, sha 5e0c3e3b4645)
- Data as of 2026-08-30T08:39:29.467469+00:00.