Skip to content

TRACK2_OMLX.md

Latest commit

 

History

History
57 lines (40 loc) · 2.45 KB

TRACK2_OMLX.md

File metadata and controls

57 lines (40 loc) · 2.45 KB
Raw

Track 2: oMLX / Claude Code wire format

Two parallel tracks share the same sandbox and curation, but different training/generation wire formats:

| Track | Training actions after <channel|> | Use case | |-------|----------------------------------------|----------| | Dingo (default) | write_file: {...} \| python: ... | Dingo web UI, JSON agent gen | | oMLX | <|tool_call>call:Read{file_path: <|"|>...<|"|>}<tool_call|> | Claude Code VS Code + oMLX |

You do not need a separate repo or host — only a convert step and/or wire_format=omlx_claude on generate/train.

In the DingoAI web UI

  1. Set Training track to oMLX / Claude Code (default stays Dingo — nothing is overwritten).
  2. Optional: Build oMLX pack from Dingo curated — writes only all_omlx_tool_training.jsonl.
  3. Train uses the oMLX pack path and passes --wire-format omlx_claude to the job.

Fast path (CLI)

python3 tools/curate_all.py
python3 tools/build_omlx_training_pack.py
./run_train_only.sh data/curated/all_omlx_tool_training.jsonl 520 models/mlx_self_training/pilot_v4_omlx "" omlx_claude

Fuse adapter → point ~/.claude/settings.json Opus at the new fused folder.

New trajectories (native oMLX format)

  1. In Dingo UI or config/default_config.json, set: 
    "generation": { "wire_format": "omlx_claude", ... }
  2. Use preset omlx-claude-prompt (tasks say Read/Write/Edit/Bash) or omlx-claude-combined.
  3. Generate as usual (run_generate.sh / web UI).
  4. Curate + merge into your pack; optional build_omlx_training_pack.py if rows are still Dingo-shaped from old runs.
  5. Train with --wire-format omlx_claude.

What was added

  • mlx_foundation/src/agent_wire_formats.py — convert + parse Gemma call:Tool{args}
  • tools/build_omlx_training_pack.py — Dingo pack → all_omlx_tool_training.jsonl
  • Sandbox: Read/Write/Edit/Bash aliases (edit, bash)
  • Generator + trainer: wire_format flag

Preset only?

A preset alone is not enough. Presets only change task wording. Track 2 also needs:

  • Training on call:Read{...} completions (--wire-format omlx_claude), and/or
  • Generation with wire_format: omlx_claude so new trajectories are native.

Claude Code after training

  • Keep using oMLX (no custom VS Code host).
  • Lower temperature on fused model in oMLX model settings (~0.3).
  • Consider disabling thinking budget for coding on that model.