doirty/haramusic-runs

HaramMusic experiments and progress archive

This repository is the canonical experiment archive for the general music remover. It stores run outputs and history; it is not the training-data repository.

Data repositories and experiment archive

• Legacy data: doirty/HaramMusic_Dataset — retain on HF, clearly mark as an archived mixed/16 kHz mono corpus, and do not use as the active background-music training set.
• New active data: doirty/haramusic-training-data — planned fresh curated corpus centered on real background/underscore music under speech, with rights/attribution manifest, quality tiers, source-disjoint splits, deterministic mixes, and QA. It has not yet been created or populated.
• Experiments: doirty/haramusic-runs — this repository, containing code snapshots, progress state, checkpoints, metrics, logs, and input/output audio. Keep it separate from both music-data repositories.

Current completed experiments

• RUNS/tiger_ft/tiger-ft-kaggle-v5-20261026/ — full v5 run artifacts.
• RUNS/tiger_ft/tiger-ft-kaggle-v6-20261026/ — full v6 artifacts, including the user listening review. The user's audition found the music removed but voice drastically quieter; v6 is not accepted as a balanced model.
• CODE/tiger_ft/v6/ — exact pushed v6 code snapshot.
• CODE/tiger_ft/v7_candidate/ — metadata-only candidate; it was not run on Kaggle.

Current state and plan

• Latest state snapshot: STATE/20261026_user_choices/PROJECT_STATE.md.
• Current dataset/model roadmap: STATE/20261026_user_choices/DATASET_AND_MODEL_LADDER_PLAN.md.
• Read-only HF checkpoint inventory: STATE/20261026_user_choices/HF_MODEL_AUDIT.md.
• Cleanup record: STATE/20261026_user_choices/HF_CLEANUP_CANDIDATES.md documents the three approved ZIP removals.
• Rights filter: CC0 or plain CC BY only, with attribution/provenance; exclude NC/ND/SA, Sampling+, research-only and unclear terms.
• The user wants the old raw corpus kept as legacy, a fresh background-music-focused dataset built, and the experiment archive separate. Model/checkpoint cleanup candidates are pending an exact dry-run list and approval; never touch unrelated projects.
• Task labels remain MUSIC and KEEP; speech/SFX/ambience/human and music-context groups remain metadata and are evaluated separately.
• Model comparison target: TIGER before/after fine-tuning plus custom approximately 5.64M, 1.17M, 293k, and 50k stereo models.

Legacy paths retained

Older lowercase code/ and overnight/ paths are historical and remain available. With the user's approval, only three superseded handoff ZIP copies were removed from the active tree; the current ZIP is STATE/20261026_user_choices/handoff_bundle.zip. No model checkpoints, run evidence, dataset audio, or unrelated project files were deleted.

Training data is separate

The public doirty/HaramMusic_Dataset is an existing mixed 16 kHz mono source corpus and will remain as a legacy archive, not the active training source. The actual v5/v6 runs used the narrower doirty/haramusic-pool-v3 Kaggle input; it was not curated specifically for background/underscore music under speech. Neither should be treated as the new canonical corpus. The new data repo is pending licensed-source curation, background-music selection, quality-tier evaluation, and per-item rights checks.

Run artifacts are post-run and preserved. The only recent HF deletions were the three approved superseded handoff ZIP copies documented in STATE/20261026_user_choices/HF_CLEANUP_CANDIDATES.md; the latest handoff remains. Check each run folder's summary.json, per-step metrics, checkpoints, and listening files before making comparisons.