MTP2 Audio Tool RL

MTP2 Audio Tool RL is a cleaned research-code repository for tool-augmented audio-language model experiments. It collects extracted project-owned code for GRPO-style training, DeSTA/Audio-Maestro/ToolRL integrations, dataset construction helpers, inference utilities, evaluation scripts, and patch-preservation metadata from a larger research workspace.

Current Repository Status

This is a public research repository, not a polished reproduction package. Lightweight setup checks, manifest validation, CI, submodule pinning, and patch dry-runs are in place. Full end-to-end training reproduction, dataset reconstruction, and heavyweight test coverage are still incomplete.

What Is Included

• Python package code under src/mtp2_audio_tool_rl/
• Run scripts and setup helpers under scripts/
• Selected configs under configs/
• Tests and diagnostics that were safe to extract
• Dataset skeleton files under manifests/
• Preserved patch metadata under patches/
• Pinned third-party submodules under third_party/

What Is Not Included

• Raw datasets or copied audio
• Model checkpoints, weights, embeddings, or caches
• Generated tool outputs, logs, or results
• Applied third-party patch changes inside submodules
• A complete paper-style reproducibility package

Repository Layout

mtp2-audio-tool-rl/
├── configs/
├── docs/
├── manifests/
├── patches/
├── scripts/
├── src/
├── tests/
└── third_party/

Documentation Map

• Quickstart
• Reproducibility
• Dataset reconstruction
• Public documentation map
• Internal audit and migration docs

Quickstart

Clone with submodules and install the lightweight base environment:

git clone --recurse-submodules https://github.com/DakshGoyal1506/mtp2_audio_tool_rl.git
cd mtp2_audio_tool_rl
python3 -m pip install -e .
python3 -m pip install -r requirements/base.txt

Run the lightweight checks:

python3 scripts/setup/check_repo_setup.py
python3 scripts/setup/check_imports.py
python3 scripts/data/validate_manifest.py manifests/example_manifest.jsonl
bash scripts/setup/apply_patches_preview.sh

More detail: docs/quickstart.md

Submodules

This repository uses pinned submodules for upstream dependencies such as Audio-Maestro, DeSTA2.5-Audio, and ToolRL. They are checked out for reference and patch dry-run validation only. The repository does not claim that those third-party components are relicensed under the top-level MIT license.

Setup Checks

The current lightweight validation surface includes:

• repository safety and artifact checks
• package import checks without heavy ML/audio dependencies
• example manifest validation
• patch dry-runs against pinned submodules

These same checks are also used in CI. Full pytest is intentionally not part of the default path yet.

Dataset Policy

Datasets stay outside Git. This repository includes only lightweight dataset metadata such as schemas, tiny fake/example manifests, and reconstruction notes. See manifests/README.md and docs/dataset_reconstruction.md.

Patch Preview Workflow

Patch files under patches/ preserve local tracked modifications from earlier working copies. The repository provides a dry-run-only preview script:

bash scripts/setup/apply_patches_preview.sh

This checks whether selected preserved patches apply cleanly to the pinned submodules, but it does not apply them.

License And Third-Party Notice

The top-level LICENSE is MIT for original repository-authored code, scripts, configuration, documentation, and scaffolding. Third-party submodules and patch-derived content retain their own terms or unresolved status. See THIRD_PARTY_NOTICES.md.

Current Limitations

• Full training reproduction is not turnkey yet.
• Datasets, raw audio, checkpoints, and generated outputs are not included.
• Heavy tests and model-dependent scripts require external assets and optional dependencies.
• Patch overlays are preserved and validated, but not applied automatically.

Reproducibility notes: docs/reproducibility.md