IndoML 2026 Datathon, Track 1: Team Arcturus checkpoints
Checkpoints of Team Arcturus (Codabench leaderboard name asymptoticchase) for Track 1 (noise event detection) of the IndoML 2026 Datathon. The system predicts the onset and offset of every noise event in Vaani field recordings.
Phase 2 result: Combined 1.5893 (Event F1 0.7691, Segment Dice 0.8202), 8th place.
- Code (training and inference): https://github.com/Abhilash1602/IndoML2026-Track1-Arcturus
- Technical report:
docs/Arcturus_track_1_8.pdfin the GitHub repository
Contents
The final system loads these 20 files (full sha256 in CHECKPOINTS.md):
| Role | File | Size (MB) | sha256 (prefix) |
|---|---|---|---|
| member va_atst | checkpoints/dualbranch_atst_frame_stage2_i2_owbceVS_atst_teacher.pth |
352.4 | 1f374bba02555d48… |
| member va_owB | checkpoints/standalone_dasheng_stage2_i2_owbceVS_owB_teacher.pth |
350.2 | c26207ac3da6a29a… |
| member va_bS2 | checkpoints/standalone_beats_stage2_owbceVS_bS2_teacher.pth |
369.7 | d79f4dbbcc53d63c… |
| member va_owBt | checkpoints/standalone_dasheng_stage2_i2_owbceVS_owBt_teacher.pth |
350.2 | e541a327b8631949… |
| member va_bA | checkpoints/standalone_beats_stage2_owbceVS_bA_teacher.pth |
369.7 | e2edfe4dd096bb45… |
| member atst_stage2 | checkpoints/dualbranch_atst_frame_stage2.pth |
352.3 | 245ea03e98d9a8c8… |
| member dasheng_owB | checkpoints/standalone_dasheng_stage2_i2_owbceB.pth |
350.2 | bba863c849231c8f… |
| member dasheng_owB_t | checkpoints/standalone_dasheng_stage2_i2_owbceB_teacher.pth |
350.2 | 34edcc06bb0cbf83… |
| member beats_s2 | checkpoints/standalone_beats_stage2.pth |
369.7 | 4d252a5abdc0c4a7… |
| member beats_owA | checkpoints/standalone_beats_stage2_owbceA.pth |
369.7 | 2c67b0f7b9cfccc2… |
| boundary refiner (voter 0) | checkpoints/refiner_val_all_x0.pth |
3.2 | 599b333ef8241ca5… |
| boundary refiner (voter 1) | checkpoints/refiner_val_train70_x0.pth |
3.2 | 45d5fe77c924c00f… |
| boundary refiner (voter 2) | checkpoints/refiner_val_all_x0_s1.pth |
3.2 | b8ba2266d90d4d27… |
| boundary refiner (voter 3) | checkpoints/refiner_val_all_x0_s2.pth |
3.2 | 73595bbbdaf9e479… |
| boundary refiner (voter 4) | checkpoints/refiner_val_all_x0_s3.pth |
3.2 | ed8a7281b81e783d… |
| FP filter + verifier (GBMs, PCAs) | checkpoints/postproc_v11.pkl |
4.2 | 13d63d852c66f504… |
| DETR-style detector (verifier evidence) | checkpoints/sedt_gold_anc_s0_alone.pth |
24.5 | d0b9f24bf3701169… |
| gated edge-correction table | checkpoints/edge_correction_v5.json |
0.0 | d6f397f15cd3f625… |
| backbone init (ATST-F) | checkpoints/atst_frame/ATST-F_strong_1.pt |
344.5 | fbf2577958e3648d… |
| backbone init (BEATs) | checkpoints/beats/BEATs_iter3_plus_AS2M.pt |
361.5 | d43cbfad4d7b5638… |
mispeech/dasheng-base is downloaded from Hugging Face automatically and is not stored here.
Usage
git clone https://github.com/Abhilash1602/IndoML2026-Track1-Arcturus && cd IndoML2026-Track1-Arcturus
pip install -r requirements_inference.txt
bash scripts/download_checkpoints.sh hf # downloads this repository into ./checkpoints and verifies sha256
python inference/infer.py <input_audio_dir> <output_dir>
System
- Frame ensemble: ten detectors on a 10 ms grid (2 × ATST-Frame with a 10 ms CNN branch, 4 × Dasheng-base, 4 × BEATs, each with a Conformer head).
- Event-level stage: decides which candidate events to keep or merge, using five boundary refiners as voters, gradient-boosted verifiers and a DETR-style set detector.
- Confidence guard: an event with mean probability ≥ 0.70 is never dropped, so one pipeline serves natural and synthetic clips.
Details are in the technical report.
Licence and pretrained models
Our checkpoints are released under the MIT licence. They build on these public pretrained models:
| Model | Source | Licence |
|---|---|---|
| ATST-Frame | PretrainedSED | MIT |
| BEATs | microsoft/unilm | MIT |
| Dasheng | mispeech/dasheng-base | Apache-2.0 |
Team
- Abhilash Jacob Mathew (team leader), HCLTech
- Aman Kumar, KIET GAZIABAD
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support