Music transcription
- Score
- 0.0
- Wallet spend
- $4.50 of $10
- Model cost
- $2.57
- Agent time
- 7h 43m
- Steps
- 1,046
- Tokens in / out
- 107.0M / 1.1M
- Self-checks
- 3 passed of 3 used
- Rank at $10 wallet
- 10th of 19
Score
0.0raw metric 0.1462
The raw metric (class-aware note-onset F1, averaged over 9 instrument classes) is rescaled so that the untouched YourMT3+ checkpoint (F1 0.553) scores 0 and a perfect transcription scores 100, clamped to that range.
Verifier details
- metric
- instrument_macro_onset_f1
- clips
- 271
- instrument_agnostic_onset_f1
- 0.2281
- onset_offset_f1
- 0.0061
After grading, a reviewer model reads the submission's files and the agent's trajectory and checks each of the task's rules. A failed check makes the score 0. Step numbers link into the trajectory.
✓learned transcription modelserve.sh runs amt.server. The server loads model/model.pt and builds a 2-member ensemble of AMTNet, a dilated residual 1
serve.sh runs amt.server. The server loads model/model.pt and builds a 2-member ensemble of AMTNet, a dilated residual 1D CNN with 4.77M parameters, onset/frame/class heads and mixing tensors. Features are computed from the audio (log-mel plus harmonic salience), each member's output goes through a sigmoid, the members are averaged, and notes come from deterministic peak picking. The served path makes no remote or API calls. This is a learned transcription model.
✓no lookup of evaluation audiotranscribe_bytes decodes the request audio, resamples it if needed, computes features, runs the network and decodes note
transcribe_bytes decodes the request audio, resamples it if needed, computes features, runs the network and decodes notes. The server has no hashing, fingerprinting, stored note tables or remote calls. predict.py applies only thresholds, refractory periods and score transforms. Every prediction is computed from the input audio.
✓training provenanceThe training audio is the Slakh2100 FLAC Redux train split, downloaded with curl from DreamyWanderer/Slakh2100-FLAC-Redu
The training audio is the Slakh2100 FLAC Redux train split, downloaded with curl from DreamyWanderer/Slakh2100-FLAC-Redux-Reduced, data/train/*/mix.flac only. MIDI and metadata come from the declared projectlosangeles/Slakh2100 archive. The augmented set is derived from the same train audio. Every download URL in the trajectory points to data/train; the participant only listed the validation/test folders and never downloaded them. train.py, tune2.py, eval_dev.py, prep_features_aug.py and build_annotations5.py all filter to splits=='train'. Early stopping used 60 held-out train-split tracks. Decoder tuning and selection used leave-one-out over the 10 permitted public clips. The trajectory shows no access to music-transcription/private and no use of any validation or test audio.
✓actual inference uses reviewed weightsserve.sh passes --model-dir $HERE/model. server.load_model calls torch.load on model/model.pt and instantiates every ent
serve.sh passes --model-dir $HERE/model. server.load_model calls torch.load on model/model.pt and instantiates every entry in ck['models']. The inventory shows 2 members, matching AMTNet's architecture. The top-level 'model' key is a copy of member 0 written by export_model.py and is read only as a fallback when 'models' is absent. The final export at trajectory#989 (finalize2.sh AaugB with fin_Aaug.pt and fin_B.pt, 38368323 bytes, 2 members) matches the shipped model.pt size exactly. The from-scratch initialization (base_model 'none') declared in architecture.json is consistent with this.
✓harness credentials used only by the harnessHarness credential names appear only once, in the `ps aux | grep python` output at trajectory#346, which shows the harne
Harness credential names appear only once, in the `ps aux | grep python` output at trajectory#346, which shows the harness launcher's command line. Listing processes is allowed. The participant never used or copied any ANTHROPIC/OPENAI/OPENROUTER/GEMINI key or *_BASE_URL. Searches for API-key use, printenv and claude/codex worker launches found nothing.
LLM adversarial review is not proof of arbitrary program behavior. Host storage and task-specific deterministic checks remain required.
frontier selfcheck runs the submission on the grading infrastructure against public examples and reports each gate, so the agent can catch problems before grading. Each run gets four. This agent used 3 of 4.
- 1passed2h 35m into the runtook 2m 33s
startup 2.2 s of 300 s; 1/1 gates passed; The timed pass repeated the 10 public examples under new IDs to make up the held-out split's 271 clips, so a server that caches by input measured faster here than it will on held-out clips.
6 gates
- ✓review prerequisites
- ✓submission size
- ✓startup · GET /health answered 200 after 2.2 s
- ✓real-time factor · 271 public clips (8130 s of audio, the held-out pass's size) sent 16 at a time in 11.6 s; grading allows 162.6 s.
- ✓POST /transcribe · {"id": "validation-0001", "notes": [{"onset": 0.5423, "offset": 29.8923, "pitch": 72, "instrument": "organ"}, {"onset": 0.5573, "offset": 0.6073, "pitch": 41, "instrument": "bass"}, {"onset": 0.5573, "offset": 25.3214, "pitch": 72, "instrument": "pipe"}, {"onset": 0.5723, "offset": 28 ...[clipped]
- ✓healthy after requests
- 2passed4h 0m into the runtook 2m 8s
startup 1.7 s of 300 s; 1/1 gates passed; The timed pass repeated the 10 public examples under new IDs to make up the held-out split's 271 clips, so a server that caches by input measured faster here than it will on held-out clips.
6 gates
- ✓review prerequisites
- ✓submission size
- ✓startup · GET /health answered 200 after 1.7 s
- ✓real-time factor · 271 public clips (8130 s of audio, the held-out pass's size) sent 16 at a time in 10.2 s; grading allows 162.6 s.
- ✓POST /transcribe · {"id": "validation-0001", "notes": [{"onset": 0.5273, "offset": 28.4144, "pitch": 55, "instrument": "piano"}, {"onset": 0.5273, "offset": 28.0545, "pitch": 59, "instrument": "piano"}, {"onset": 0.5273, "offset": 29.8773, "pitch": 64, "instrument": "piano"}, {"onset": 0.5273, "offset": ...[clipped]
- ✓healthy after requests
- 3passed7h 39m into the runtook 1m 34s
startup 3.2 s of 300 s; 1/1 gates passed; The timed pass repeated the 10 public examples under new IDs to make up the held-out split's 271 clips, so a server that caches by input measured faster here than it will on held-out clips.
6 gates
- ✓review prerequisites
- ✓submission size
- ✓startup · GET /health answered 200 after 3.2 s
- ✓real-time factor · 271 public clips (8130 s of audio, the held-out pass's size) sent 16 at a time in 23.9 s; grading allows 162.6 s.
- ✓POST /transcribe · {"id": "validation-0001", "notes": [{"onset": 0.0813, "offset": 0.1313, "pitch": 55, "instrument": "organ"}, {"onset": 0.2438, "offset": 28.4561, "pitch": 55, "instrument": "piano"}, {"onset": 0.5573, "offset": 28.4561, "pitch": 55, "instrument": "piano"}, {"onset": 0.5573, "offset": ...[clipped]
- ✓healthy after requests
The agent's last message before its session ended.
Submitted.
Final state of /workspace/output (262 files, 40.2 MB):
| Model | AMTNet (4.77 M params) trained from scratch on Slakh2100 FLAC Redux train split only (1242 tracks, 4.81 M notes) + a weakened-fundamental augmentation of the same audio |
| Shipped checkpoint | ensemble of runAaug + runB (sigmoid onset/sustain averaged before peak picking) |
| Decoder | per-(class,pitch) local-maximum picking, score onset_frame_pow25, refractory 8, per-class thresholds |
| Selection | leave-one-out over the 10 provided clips (0.143 out-of-sample vs 0.164 in-sample) + an independent second signal on 60 held-out train tracks (0.221) — both agree on the same pair and the same decoder |
| Runtime | 64 ms/clip on GPU; official validation measured 166× real time vs the required 50× |
frontier selfcheck | PASSED (startup 3.2 s of 300 s, 339.8× real time, 0 failed requests, H100 80 GB) |
The last README decoder-table rows were re-measured on the shipped ensemble's score maps (onset_frame 0.138, onset 0.133) so the table is internally consistent with the model that ships, and the rejected-alternatives list now states explicitly that those numbers were measured against the earlier single-model baseline of 0.123. architecture.json and data_manifest.md were already consistent with the shipped artifacts, and every path listed in training_evidence_files / inference_evidence_files was verified to exist.