localrl

KO | EN

gitlite — search

#ai #large-language-models #shell #web #editor #nextjs #claude-code #android #video #openai

(★ 9)

Local domain RL on Apple Silicon — train in the browser with GRPO.

File Explorer

ZIP Download

- - training.yaml
- - corpus.jsonl
  - dataset.json
  - eval.jsonl
  - probe.jsonl
  - train.jsonl
- - corpus.jsonl
  - dataset.json
  - eval.jsonl
  - probe.jsonl
  - train.jsonl
- - pipeline-scoreboard.jpg
  - playground-compare.jpg
- hero.png
- __init__.py
- compare.py
- config.py
- hwinfo.py
- inference.py
- model_catalog.py
- runstore.py
- tool_call_compat.py
- tool_format.py
- tool_stream_parser.py
- - __init__.py
  - retrieval_qa.py
- __init__.py
- domain.py
- evaluate.py
- gate.py
- grpo.py
- pipeline.py
- rollout.py
- sft.py
- train.py
- __init__.py
- anchors.py
- bundle.py
- data.py
- decoding.py
- env.py
- evaluate.py
- gate.py
- gen_dataset.py
- gen_repo_dataset.py
- grpo.py
- memprobe.py
- patch.py
- rollout.py
- run_all.py
- scoring.py
- sft.py
- tasks.py
- tools.py
- train.py
- __init__.py
- bench_rollout.py
- build_pubmedqa.py
- - index.html
  - install.md
- __init__.py
- app.py
- __init__.py
- export_tasks.py
- generate.py
- importer.py
- judge.py
- quality.py
- rl_pool.py
- runner.py
- scoring.py
- tasks.py
- test_anchors_and_advantage.py
- test_build_pubmedqa.py
- test_compare_domain_serving.py
- test_compare_record_gold.py
- test_degeneracy_penalty.py
- test_direct_task_reward.py
- test_domain_launch_guard.py
- test_domain_macro_metrics.py
- test_domain_rl.py
- test_domain_rollout_framing.py
- test_entropy_bonus.py
- test_env_robustness.py
- test_evot_anthropic_bridge.py
- test_failed_adapter_release.py
- test_final_answer_penalty.py
- test_force_answer.py
- test_force_landing.py
- test_gate_release.py
- test_landing_penalty.py
- test_loop_and_exit_signal.py
- test_memory_and_exploration.py
- test_model_catalog_attribution.py
- test_pipeline_cancel.py
- test_readonly.py
- test_repeat_penalty.py
- test_reward_explanation.py
- test_reward_kinds.py
- test_rollout_concurrency_equiv.py
- test_rollout_invalid_tool.py
- test_sft_staging.py
- test_skip_sft_wiring.py
- test_tool_call_compat.py
- test_tool_format.py
- test_tool_stream_parser.py
- test_trajectory_monitor.py
.gitignore
AGENTS.md
install.md
pytest.ini
README.md
requirements.txt

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

![LocalRL](https://raw.githubusercontent.com/evotai/localrl/main/docs/hero.png)

**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.

```text
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

### What a ~20 min local run learns

Reference: [`pipe-20260716-213411`](https://rl.evot.ai/#pipe/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:

| | Base | SFT | **RL** |
|---|---:|---:|---:|
| macro grounded accuracy | 0% | 63.3% | **66.7%** |
| exact quote | 0% | 96.7% | **98.3%** |
| unsupported claims | — | 1.7% | **0%** |

Playground side-by-side ([cmp](https://rl.evot.ai/#playground/cmp-20260716-220016)) 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.

## Live preview

**https://rl.evot.ai**

[Pipeline](https://rl.evot.ai/#pipe/pipe-20260716-213411)

<sub>scoreboard · stages · progression</sub>

</td>
<td width="50%" valign="top" align="center">

[Playground](https://rl.evot.ai/#playground/cmp-20260716-220016)

<sub>base vs RL · tool calls · interaction flow</sub>

</td>
</tr>
</table>

## Run

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:

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

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

<details>
<summary>Manual shell setup</summary>

```bash
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
```

</details>

## Datasets

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.

## API

```text
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.

localrl

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

localrl

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README