---
license: apache-2.0
base_model: mlx-community/Qwen3-1.7B-4bit
tags:
- mlx
- lora
- doer
- qwen3
- unix
library_name: mlx
---

# cagataydev/doer

The **default checkpoint** for [doer](https://github.com/cagataycali/doer) —
a one-file pipe-native self-aware Unix agent.

## what

A LoRA-fine-tuned `mlx-community/Qwen3-1.7B-4bit` that knows:

- what doer is, its architecture, its SOUL (creed)
- all `DOER_*` env vars and their defaults
- how to train, upload, round-trip data via `--train*` / `--upload-hf`
- the design rules: one file, lean deps, context over memory, unix over RPC,
  env vars over config files
- how to use doer with images, audio, video (mlx-vlm routing)
- provider auto-detection (bedrock → mlx → ollama)

## use

```bash
pip install 'doer-cli[mlx]'

# point at this checkpoint
DOER_PROVIDER=mlx \
DOER_MLX_MODEL=cagataydev/doer \
doer "what is doer"
```

Future doer builds default `DOER_MLX_MODEL=cagataydev/doer`, so:

```bash
pip install 'doer-cli[mlx]'
doer "what is doer"   # auto-pulls this checkpoint on first run
```

## training

- **base**: `mlx-community/Qwen3-1.7B-4bit`
- **data**: [cagataydev/doer-training](https://huggingface.co/datasets/cagataydev/doer-training)
  (fat, self-contained records: `{ts, query, system, messages, tools}`)
- **method**: LoRA via `mlx_lm.tuner`, 8 layers, rank 8, scale 20
- **fused**: `mlx_lm.fuse --dequantize` → re-quantized to 4bit

Trained on self-generated Q/A turns about doer itself — the model learns its
own source, its own prompt, its own philosophy.