MaximeM/Qwen3-Embedding-4B-recsys-challenge-2026-retriever-ctx1024

Qwen3-Embedding-4B-recsys-challenge-2026-retriever-ctx1024

Dense dual-encoder retriever for conversational music recommendation, from team FPMs_UMONS's entry to the RecSys Challenge 2026. It is the dense channel of the four-channel retrieval pool (BM25, dense, item-based collaborative filtering, artist expansion) that feeds our scoring-head reranker.

Fine-tuned from Qwen/Qwen3-Embedding-4B.

This is a pipeline component, not a standalone model. It is one channel of a retrieval-and-reranking system and is only meaningful inside it. There is no standalone usage recipe here on purpose: how to download it, what text to encode, and how its output is fused with the other three channels are all documented — and executable end to end — in the repository.

• —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.
• —Reranker it feeds: `MaximeM/Qwen3-8B-recsys-challenge-2026-scorehead-top200`

Retrieval recall figures and ablations are reported in the paper and the repository; they are not duplicated here.

Critical: context length and truncation side

The model must be used with max_seq_length = 1024 and truncation_side = 'left', matching training. The current user request is placed last in the query text, and most dev queries exceed 256 tokens — with default right truncation the request is silently thrown away on every long conversation, which costs most of the model's value. The pipeline sets both; anything loading these weights directly has to do the same.

Training

LoRA on attention, MultipleNegativesRankingLoss with cross-device gather, no hard negatives, at sequence length 1024 with left truncation. Merged after training and exported in sentence-transformers format.

Limitations

• —English, TalkPlayData-style conversations only.
• —Tuned for the challenge's ~47k-track catalog: this is a domain retriever, not a general-purpose embedding model — use the base Qwen3-Embedding-4B for that.
• —Retrieval pools built with it are bit-reproducible, unlike the downstream reranker.

Licence and attribution

Released under the Apache License 2.0, inherited from Qwen/Qwen3-Embedding-4B. A copy of the licence is included as LICENSE.

Modifications: these weights are a fine-tuned derivative of Qwen3-Embedding-4B, adapted with LoRA on the attention projections and merged, with the sequence length and truncation side changed as described above.