DingoAI

Train a local Gemma model to use tools the way Claude Code does — write_file, read_file, list_dir, and python in a sandbox — without sending data to a cloud API.

You point teacher models at tasks, they produce multi-step trajectories, the sandbox checks that tests actually pass, and MLX LoRA fine-tunes your student on the ones that work. Everything runs on your Mac.

Repo: github.com/True2456/DingoAI

---

What you get

1. Generate — teachers write code and run unittest in an empty sandbox until a trajectory passes.
2. Curate — merge runs, drop weak rows, build data/curated/all_tool_training.jsonl.
3. Train — LoRA on the student (default: Gemma 4 26B MoE).

There is a second oMLX / Claude Code wire format (call:Read{…} style) for serving through oMLX. See docs/TRACK2_OMLX.md.

---

Quick start

git clone https://github.com/True2456/DingoAI.git
cd DingoAI
./run_web_gui.sh

Opens http://127.0.0.1:8765 on this machine only.

What	Command
Web UI	./run_web_gui.sh
Smoke test	./run_smoke.sh
Generate 100 samples	./run_generate.sh 100
Train on curated data	./run_train_only.sh data/curated/all_tool_training.jsonl 120
Rebuild curated pack	python3 tools/curate_all.py

Weights live under ~/.lmstudio/models/ (or paths you set in the console). They are not in this repo.

---

Web console

The UI handles model paths, prompt presets, generate/train jobs, and live logs. Useful bits:

• Presets in config/dingo_presets.json (Antigravity-style security, red team, networking, JSON-patch focus, oMLX track).
• Training track — Dingo tools vs oMLX Claude markup.
• Build oMLX pack — writes data/curated/all_omlx_tool_training.jsonl from the Dingo curated file.
• Train ratio warning when iterations ÷ samples gets too high.

Regenerate the screenshot:

pip install playwright && playwright install chromium
python3 tools/capture_gui_screenshots.py

---

Numbers from local runs

data/ is gitignored; these come from JSONL on disk as of May 2026.

What	Count	Notes
Merged curated pack	349	data/curated/all_tool_training.jsonl after curate_all.py
Saved trajectories (all runs)	454	Lines across data/generated/*.jsonl (not counting *_failed_attempts.jsonl)
qwencoder7 batch	5 saved, 23 failed attempts	Hard Antigravity-style tasks; most failures were bad JSON from teachers, not sandbox rejects — details
Read-before-patch runs	30/30 saved → 28 kept	NewModelRun7-ReadFocus JSONL; curation dropped 2
V3_Jsonpatch2	36 saved → 32 kept	After sort/curation

Early batches (generic tooling prompts, Qwen-first teacher order) often saved well under half of requested samples. After read_file-focused prompts and putting Gemma first in the teacher order, 30-sample runs commonly save all lines to JSONL; curation may still trim a few.

Do not read *_failed_attempts.jsonl line counts as “samples requested” — each line is usually one teacher attempt that did not produce a kept trajectory.

---

How generation fails (and what we changed)

Typical failure modes, in order of how often they showed up in logs:

1. Teacher returns malformed JSON → turn discarded.
2. Patch task finishes without a second python after the fix → workflow reject.
3. Task wording ambiguous (e.g. “reverse words” interpreted differently by different models).

Five reference tasks were run in Antigravity (Gemini) and compared to MLX teachers on the same instructions. MLX saved 0/5 on that set in qwencoder7; Antigravity completed all five. Write-ups: docs/generation_findings.md.

Prompt rules learned from that work are baked into mlx_foundation/src/generator/generator.py (fail-then-patch flow, unittest only, read_file before patch when required).

---

Layout

Path	Purpose
mlx_foundation/src/generator/	Task bootstrap + multi-teacher trajectories
mlx_foundation/src/sandbox/	Tool execution and test verification
mlx_foundation/src/trainer/	LoRA training
web/	Local console
tools/curate_*.py	Tier, merge, promote false rejects
config/dingo_presets.json	Named prompt + model presets

---

Training notes

On MoE students, keep training iterations ÷ curated samples ≤ ~3× to limit memorization. The console shows the ratio when you pick a JSONL file.

LoRA is self-attention only (rank 16, LR 1.5e-6) so routing stays stable.

---

Two Macs

Generate on a second machine, copy the JSONL, train on the main one:

Second Mac:  ./run_generate.sh 100  →  batch.jsonl
Primary Mac: ./run_train_only.sh batch.jsonl

---

License

MIT