Graded chunk relevance distilled from a query-aware context compressor, no re-labeling needed
antoine_chaffin · x · 2026-11-03
The approach distills graded chunk-level relevance from a query-aware context compressor. The teacher scores every token, and those scores are aggregated over arbitrary chunk boundaries — giving richer supervision that works with any chunking strategy without re-labeling documents.
The attached figure contrasts labeling schemes: the usual label marks one chunk relevant and all others irrelevant, while pplx-emvbed-v2-context scores every chunk (highest for the answer chunk, medium for context chunks, near zero for irrelevant ones). Scores come from the context compression model in Perplexity's answer pipeline; chunk scores form a distribution over the document's chunks that the model learns to match, plus the usual document-level contrastive loss.
Related event: Distilling Graded Chunk Relevance Fixes Contextual Embedding Training(2 posts)→
More from Models
- Custom fine-tunes now beat frontier models at ~1/20th the cost, says investor — bindureddy · 2026-11-03
- OpenAI updates Model Spec: intent over literal instruction-following — sven_ai · 2026-11-03
- Zyphra's ZHA brings subquadratic attention via hierarchical tree-structured indexer — yacineMTB · 2026-11-03
- LiquidAI's LFM3-1.7B tops HF trending with 1k likes and 42.4k downloads in 5 days — maximelabonne · 2026-11-03
- 'Gemini is your friend and alignment is complicated': another eyebrow-raising model reply — vedantmisra · 2026-11-03
- Pedro Domingos: Humans are much more resistant to data than AI is — pmddomingos · 2026-11-03