CrossQ reproduction bundle

This bundle contains a deterministic toy proxy for CrossQ: Task-Aligned Cross-Token Conditional Quantization for Late Interaction Retrieval (OpenReview: https://openreview.net/forum?id=jxHF6z0S1L).

What was reproduced

• Mechanism: synthetic document token vectors are scored with late interaction; token residuals are quantized conditionally on a document mean context computed at indexing time.
• Local run: 120 documents, 48 queries, 8 tokens/document, 32 dimensions, 2-bit codes, seed 4231.
• Results: full-precision MRR@10 proxy 0.510417; global 2-bit proxy 0.520833; conditional proxy 0.500000; conditional index compression 9.846x.

This is explicitly a toy reproduction. The checkout contained no CrossQ implementation, checkpoint, benchmark dataset, or evaluation script, so the paper's reported 61x compression and 2.3% MRR@10 loss were not independently verified. A Hugging Face GPU Job was also blocked before launch by an HTTP 402 insufficient-credit response.

Contents

• toy_crossq.py — self-contained NumPy experiment.
• outputs/toy_results.json — recorded metrics.
• poster.html and poster_embed.html — Posterly reproduction poster.

Rerun

python toy_crossq.py --output outputs/toy_results.json

The published logbook is: https://huggingface.co/spaces/phantomcoder1996/repro-crossq-task-aligned-cross-token-conditional-quantization-for-late-interaction-retrieval