See domain RL happen on your Mac — in a browser, with every step open.

LocalRL is a hands-on training dashboard for Apple Silicon. You pick a domain dataset, run SFT → GRPO → eval → release gate from the UI, and watch full traces of what the model actually did: tool calls, observations, rewards, and answers. Not a black-box CLI job — a product you can open, click through, and understand.

dataset pack → base eval → SFT → local GRPO → RL eval → promotion gate → local API

• UI-first — launch and inspect training in the dashboard, not a pile of scripts
• Full I/O traces — every SFT demo and RL rollout is expandable (tools, evidence, grades)
• Local & small — MLX on Apple Silicon; a 0.6B evidence-agent run finishes in ~20 min

Reference: pipe-20260716-213411 · Qwen3-0.6B-4bit + pubmedqa-small on M5 Pro / 24 GB (~20 min wall, peak ~6.4 GB).

On the held-out probe, RL is released over SFT:

Playground side-by-side (cmp) on “Does a dedicated discharge coordinator improve the quality of hospital discharge?”:

• Base — no tool use · invalid schema ("no|maybe") · wrong
• RL — search_papers → real PubMed id + exact quote → yes correct

Same 0.6B weights: frozen base cannot follow the evidence contract; after local SFT+GRPO the model retrieves, cites, and answers — and you can open the run to see every tool I/O that got it there.

https://rl.evot.ai

Pipeline scoreboard · stages · progression

Playground base vs RL · tool calls · interaction flow

Requires Apple Silicon Mac (16 GB+ unified memory recommended), Python 3.10+, Git.

Paste this into a coding agent — it installs and starts the dashboard for you:

https://rl.evot.ai/install.md

Then open http://localhost:7600. Default model: Qwen3-0.6B-4bit. Default pack: datasets/pubmedqa-small.

git clone https://github.com/evotai/localrl
cd localrl
python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt
python -m server.app

Default pack: datasets/pubmedqa-small (PubMedQA biomedical evidence QA).

It trains a local evidence agent — search abstracts, cite an exact quote, answer or abstain when evidence is missing. Other domain packs can be added under datasets/ and selected in the UI.

OpenAI     http://localhost:7600/openai/v1
Anthropic  http://localhost:7600/anthropic/v1

Released domain models run their local environment before answering. Keep the dashboard on 127.0.0.1 when it can launch training jobs.