# Baseline Continuations from GT Context

This directory contains 30 MP3/MIDI pairs. MP3s are exported from archived lossless masters and end at the last generated note plus up to 1.5 seconds of release tail. Original WAV files, synthesis logs, and native outputs remain in the repository's `demo/baseline-20260921/` and experiment directories. WAV hashes in `manifest.json` identify archived originals; MP3 hashes identify the current recordings.

Six original baseline models generated one continuation for each of five fixed contexts. Every player includes the same eight-bar GT prelude followed by actual generated notes, capped at the requested 32-bar continuation. The actual future was withheld from the generators.

Sampling parameters and seeds are fixed, with no metric-based selection or resampling. No training or fine-tuning was performed; these are not metric-tuning variants. Ours retains its frozen recordings and original history of up to six phrases (48 bars), with short input histories padded under the original protocol. Baselines receive eight bars, with the two four-bar exceptions below. This is a qualitative display with different training settings and internal history lengths, not a controlled ranking or final paper result.

For Melody Progression (`song-2`), BEAT and Text2midi use the last four GT context bars because the full eight-bar prefix exceeds their 2,048-token limit. Initial failures and subsequent adaptation records are preserved. The listening prelude remains eight bars for every player.

Native output may end early or reach a generation limit. Short outputs are presented at their actual length with no silent padding, looping, stretching, or added notes. Cards report actual output lengths and observed limits for each run. BEAT, Music2Music-PT, and Text2midi reach 2,048 tokens in these runs. Some Amadeus-S runs reach 3,072 tokens, and some NotaGen runs reach 1,024 patches. These limits do not imply a universal maximum length in bars. ABC quantization, velocity loss, and irregular bar lengths remain documented in the records.

| Model | Context | Input bars | Last generated note ends at | Notes |
| --- | --- | ---: | ---: | ---: |
| MuPT 1.97B | song-1 · Track01646.npz:73 | 8 | 16.2 / 32 | 658 |
| MuPT 1.97B | song-2 · Track01631.npz:32 | 8 | 4.4 / 32 | 611 |
| MuPT 1.97B | song-3 · Track01672.npz:16 | 8 | 13.1 / 32 | 1003 |
| MuPT 1.97B | song-4 · Track01657.npz:25 | 8 | 15.6 / 32 | 545 |
| MuPT 1.97B | song-5 · Track01527.npz:112 | 8 | 9.0 / 32 | 1241 |
| NotaGen RL3 | song-1 · Track01646.npz:73 | 8 | 23.2 / 32 | 1391 |
| NotaGen RL3 | song-2 · Track01631.npz:32 | 8 | 18.0 / 32 | 1395 |
| NotaGen RL3 | song-3 · Track01672.npz:16 | 8 | 32.0 / 32 | 2509 |
| NotaGen RL3 | song-4 · Track01657.npz:25 | 8 | 31.4 / 32 | 937 |
| NotaGen RL3 | song-5 · Track01527.npz:112 | 8 | 31.4 / 32 | 1709 |
| Amadeus-S | song-1 · Track01646.npz:73 | 8 | 32.0 / 32 | 1790 |
| Amadeus-S | song-2 · Track01631.npz:32 | 8 | 26.0 / 32 | 2607 |
| Amadeus-S | song-3 · Track01672.npz:16 | 8 | 27.4 / 32 | 2660 |
| Amadeus-S | song-4 · Track01657.npz:25 | 8 | 32.0 / 32 | 1118 |
| Amadeus-S | song-5 · Track01527.npz:112 | 8 | 32.0 / 32 | 1411 |
| BEAT | song-1 · Track01646.npz:73 | 8 | 5.8 / 32 | 252 |
| BEAT | song-2 · Track01631.npz:32 | 4 | 3.2 / 32 | 303 |
| BEAT | song-3 · Track01672.npz:16 | 8 | 3.2 / 32 | 141 |
| BEAT | song-4 · Track01657.npz:25 | 8 | 6.0 / 32 | 176 |
| BEAT | song-5 · Track01527.npz:112 | 8 | 3.2 / 32 | 194 |
| Music2Music-PT | song-1 · Track01646.npz:73 | 8 | 10.0 / 32 | 527 |
| Music2Music-PT | song-2 · Track01631.npz:32 | 8 | 5.0 / 32 | 389 |
| Music2Music-PT | song-3 · Track01672.npz:16 | 8 | 6.9 / 32 | 517 |
| Music2Music-PT | song-4 · Track01657.npz:25 | 8 | 19.1 / 32 | 616 |
| Music2Music-PT | song-5 · Track01527.npz:112 | 8 | 9.0 / 32 | 417 |
| Text2midi | song-1 · Track01646.npz:73 | 8 | 0.6 / 32 | 29 |
| Text2midi | song-2 · Track01631.npz:32 | 4 | 3.0 / 32 | 177 |
| Text2midi | song-3 · Track01672.npz:16 | 8 | 2.0 / 32 | 54 |
| Text2midi | song-4 · Track01657.npz:25 | 8 | 6.1 / 32 | 251 |
| Text2midi | song-5 · Track01527.npz:112 | 8 | 1.1 / 32 | 40 |

Rendering uses 44.1 kHz, a common SoundFont, FluidSynth gain 0.2, post-processing gain 1.0, no reverb/chorus, and no per-clip normalization. MIDI prefixes match. The verified lossless GT prefix is reused until 0.1 seconds before handoff; the boundary and continuation are synthesized from combined MIDI. Archived WAV prefixes match sample by sample. Current MP3s use 192 kbps and retain up to 1.5 seconds of release tail. MIDI notes, tempo, and gain are unchanged.

Inference and failure records: `experiments/cross_model/baseline_listening_20260921`. Checkpoint paths, SHA-256 hashes, full inference parameters, seeds, metrics, output sources, and trimming provenance are in [manifest.json](manifest.json).

Original generation used `scripts.prepare_baseline_listening_demo`, followed by `scripts.run_cross_model_pilot` for the six models. Only overlong-input failures used the last four bars for retry, followed by `scripts.publish_baseline_listening_demo --publish`. Current MP3 endpoints follow the last exported MIDI note-off plus 1.5 seconds, capped at the source WAV duration. The historical grid measurements in the table are preserved; Text2midi on Steady Beat has a slightly earlier MIDI endpoint (6.0 bars), which is used for playback and the card label. [native_length_verification.json](native_length_verification.json) records the checked durations and hashes.