You can deploy this model and start earning money today!

MaximeM/
Qwen3-8B-recsys-challenge-2026-scorehead-top200

Name: Qwen3-8B-recsys-challenge-2026-scorehead-top200
Availability: InStock

safetensors

qwen3

recommender-systems

conversational-recommendation

Views

No views yet

Qwen3-8B-recsys-challenge-2026-scorehead-top200

Listwise scoring-head reranker for conversational music recommendation, from team FPMs_UMONS's entry to the RecSys Challenge 2026. This is the reranker behind our final Blind-B submission.

Fine-tuned from Qwen/Qwen3-8B with LoRA (rank 16), then merged.

This is a pipeline component, not a standalone model. It reranks a candidate pool produced by our retriever, with candidates rendered in one exact format and enriched with a descriptor cache — none of which it builds for itself. There is no standalone usage recipe here on purpose: the repository is what loads it correctly and runs it end to end.

Code, weights download and full instructions: https://github.com/MaximeUM/recsys-challenge-2026
All released artifacts: RecSys Challenge 2026 — FPMs_UMONS
Paper: Picking is Not Ranking, and Explanation Quality Has Many Dimensions: Lessons for Conversational Music Recommendation. Maxime Manderlier, Fabian Lecron. RecSys Challenge 2026.

All evaluation numbers, ablations and negative results are reported in the paper and the repository; they are not duplicated here.

What it does

Given a conversation and a pool of candidate tracks, the model appends a marker token after each candidate, reads that marker's hidden state in a single forward pass, and maps it to a score with a linear head trained with listwise cross-entropy. It ranks the whole pool at once rather than emitting a single pick — which is the distinction the paper is about.

Why it will not load as a normal checkpoint

Three things make these weights unusable outside the pipeline unless reproduced exactly:

config.json declares Qwen3Model — the backbone without an LM head. Loading with AutoModelForCausalLM does not give you this model.
The scoring head is not in the safetensors. It lives in score_head.pt, which also records the marker token id; fetching only the safetensors gives a silently unusable model.
Each candidate must be rendered as one line ending with the marker token (<|box_end|>, id 151649), carrying the distilled {sound | themes} descriptors the model was trained on:

12. Thunderstruck by AC/DC [rock, 80s] {sound: hard rock, metal | themes: rebellion, anger}<|box_end|>

The descriptors are not optional, and the rendering is not approximate — another format degrades the model sharply.

The repository handles all three, and the descriptor cache ships as a released artifact alongside the commands that rebuild it.

Training

Base Qwen3-8B, frozen, LoRA rank 16 (alpha 32), merged after training.
Conversations from the challenge train split, each with a top-200 candidate pool from our four-channel RRF retrieval, candidates enriched as above.
Listwise cross-entropy over the marker scores, ground truth positive.
Session-level train/validation split (all turns of a conversation on the same side) with early stopping — scoring heads overfit from epoch 2.
Context length 14336 tokens.

Limitations

English, TalkPlayData-style conversations only.
Inference is bf16 greedy and therefore not bit-reproducible on GPU.
The marker must be a single token, so community re-tokenizations of the base model are not interchangeable.

Licence and attribution

Released under the Apache License 2.0, inherited from Qwen/Qwen3-8B. A copy of the licence is included as LICENSE.

Modifications: these weights are a fine-tuned derivative of Qwen3-8B. The transformer weights were adapted with LoRA (rank 16) and merged; the original language-modelling head was dropped and replaced by the linear scoring head distributed in score_head.pt.

MaximeM/Qwen3-8B-recsys-challenge-2026-scorehead-top200