recsys2026-minimaestro-lambdarank

MiniMaestro - LambdaRank Reranker (RecSys Challenge 2026, Blind B)

The trained LightGBM LambdaRank (LambdaMART) reranker from MiniMaestro (Team Overfit & Chill), our submission to the ACM RecSys Challenge 2026 Music-CRS (TalkPlay). This is the exact model behind our graded Blind B result: composite 0.48 (nDCG@20 0.33, LLM-as-a-Judge 4.15/5).

• File: lambdarank_model_blindsetB.txt (LightGBM text format)
• Code / full pipeline: https://github.com/simransu/recsys-2026-music-crs

What it does

It reranks a pooled candidate set from ten retrieval channels using 43 source-aware features (per-channel rank and presence, artist/album anchor overlap, entity and query-text matches, and dialogue position). It replaces fixed reciprocal-rank fusion with a learned model. See the repository README for the full architecture.

Usage

# In the pipeline repo, place the model where inference reads it:
cp lambdarank_model_blindsetB.txt cache/lambdarank_model.txt

Or load directly with LightGBM:

import lightgbm as lgb
model = lgb.Booster(model_file="lambdarank_model_blindsetB.txt")

Training data & license

Trained only on the challenge-provided TalkPlay data (talkpl-ai/TalkPlayData-Challenge-*), which is licensed CC-BY-NC. Because these weights are derived from that data, they are released under CC-BY-NC-4.0 for non-commercial research and reproduction only, consistent with the challenge terms. Please attribute the TalkPlay dataset and the RecSys Challenge 2026 organizers.

Citation

MiniMaestro: Resource-Conscious Conversational Music Recommendation with a Single Open-Weight 8B Model. Simran Sundrani and Mohan Bhambhani (Team Overfit & Chill), RecSys Challenge 2026: Music-CRS (TalkPlay).