Room to Improve in Lightweight Reranker Training
tomaarsen · x · 2026-07-17
This blog post adds several key points:
- Training used only 213K training groups for a single epoch; the author believes there is still a clear ceiling compared to methods using millions of samples or distilling from a larger teacher.
- The training data is in English, but the model's French performance on ViDoRe V3 didn't drop—it actually outperformed English in three domains.
- The author candidly lists a few failed attempts:
- Tournament scheduling caused a 27 NDCG drop in metrics;
- Gains from first-token readout were capped at roughly 1.6x;
- Reading a listwise model in a pointwise manner lost most of the improvements;
- For the 2B version, the slowest part isn't decoding, but the ViT encoder.
Related event: LightOn-rerank Targets Mixed-Corpus RAG With Multimodal Rerankers(10 posts)→
More from Research
- UK AISI: under $80 of fine-tuning fully restores filtered bio capabilities in open-weight models — connoraxiotes · 2026-10-27
- RFD4-Proteina runs on an RTX 3060: protein design from $0.001 per design — AllThingsApx · 2026-10-27
- University of Freiburg's Generative Intelligence Lab recruiting funded PhDs via ELLIS, apply by Oct 31 — yuyinzhou_cs · 2026-10-27
- Nature Physics: reprogrammable refractive index enables arbitrary multimode wave control for ML — MarioKrenn6240 · 2026-10-27
- Wakaru open-sources ontologize: labeling datasets from model activations — amplifiedamp · 2026-10-27
- EMNLP Oral paper traces multilingual representations in LLMs with cross-layer transcoders — ZhijingJin · 2026-10-27