Downloads · 30 days
0
music-crs-challenge-2026
komekami/music-crs-challenge-2026
music-crs-challenge-2026 is a machine learning model from komekami. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for lightgbm.
Trained LightGBM models behind the final submission of team Komekami at RecSys Challenge 2026: Conversational Music Recommendation (Music-CRS).
Access
Public
Updated Jul 27, 2026
Repo size
—
Likes
0
Public
Repo makeup
Click a slice to open those files.
.txt10.8 MB · 100%
At a glance
- Library
- lightgbm
- Access
- Public
- Created
- Jul 27, 2026
- Updated
- Jul 27, 2026
- SHA
- eeaf6727
Model card
From the Hugging Face model README
Music-CRS 2026 (team Komekami): trained GBDT weights
Trained LightGBM models behind the final submission of team Komekami at RecSys Challenge 2026: Conversational Music Recommendation (Music-CRS).
- Code repository (pipeline, reproduction commands, docs): https://github.com/tmp-friends/music-crs-challenge-2026
- Challenge: https://nlp4musa.github.io/music-crs-challenge/
- Final result: Blind B composite 0.4998 / nDCG@20 0.3528, 16th of 40 teams overall (Industry Track: 8th), submission
submission_exp116_lgbm6_nocs_B.zip
Both stages are trained only on the official challenge datasets (talkpl-ai/TalkPlayData-Challenge-*). No external data is used. The models are fully retrainable from scratch with the commands in the code repository README; these files are published for transparency and for direct verification without retraining.
Files
| File | Role |
|---|---|
stage1_anchor/lgbm_B_ll_top500.txt | Stage-1 BM25-anchor ranker. LightGBM binary logloss over the top-500 RRF union of 4 BM25 views + 4 exact-match fields (config exp015_B_ll_t500_B.yaml, source profile B). Referenced by reranker_model_name in that config. |
stage2_reranker_lgbm6_nocs/lgbm_lambdarank_seed2026061{6..21}.txt | Stage-2 reranker of the selected final submission: 6 LightGBM LambdaRank boosters (seeds 20260616 to 20260621). Trained on the Dev split candidate union (12 sources, cross_session dropped): 2,949,717 candidate rows, 4,333 positive groups, 73 rank-based features, num_boost_round=450, num_leaves=31, min_data_in_leaf=80, lambda_l2=8.0, learning_rate=0.03, deterministic=true. Raw scores of the 6 boosters are averaged, then the top-20 per session-turn is taken. |
Feature names are embedded in each model file (LightGBM text format). The 73 features are rank-derived only (anchor rank flags, per-source rank features, aggregate hit/RRF statistics, turn position); see build_features in mcrs/experiments/exp027_wide_source_lgbm/rank_source_lgbm_mixed_topk.py of the code repository.
Usage
import lightgbm as lgb
booster = lgb.Booster(model_file="stage2_reranker_lgbm6_nocs/lgbm_lambdarank_seed20260616.txt")
scores = booster.predict(x) # x: [n_candidates, 73] built by build_features()
The end-to-end inference flow (candidate generation, feature building, 6-seed averaging, response generation) is documented in the code repository README.
Provenance and verification
Stage-2 training is deterministic (deterministic=true, fixed seeds, fixed thread count). The published boosters were regenerated with the exact final-run command path and verified against the original run artifacts tracked in the code repository:
- Dev full-fit nDCG@20 matches the original run report value
0.25870485648368946exactly. - The Blind B top-20 ranking produced by these boosters is identical, on all 80 of 80 rows, to the canonical ranking that produced the final submission (0 mismatching rows); reloaded files reproduce in-memory scores with zero difference.
Contact
Team Komekami. See the code repository for details.
- Library
- lightgbm
- Created
- Jul 27, 2026
- Updated
- Jul 27, 2026