vein05/ragscale-interaction-matrix

ragscale Interaction Matrix

Reader answers under raw and compressed RAG evidence: 176,864 rows, one per benchmark item, reader model, and evidence policy, across LongMemEval, HotpotQA, MuSiQue, and NQ-Open.

This is the interaction matrix released with the paper Compression Is Not Evaluation-Neutral: Fixed RAG Compression Can Distort Reader Comparisons (Sugam Panthi and Rabab Abdelfattah, arXiv:2606.21807). The paper gives 8 to 20 reader models the same stored compressed text and measures how much of a reader upgrade survives compression. On HotpotQA, one stored RECOMP output shrinks a 31.8-point gap between the lowest- and highest-scoring readers to 7.8 points.

• Paper: https://arxiv.org/abs/2606.21807
• Code (ragscale audit toolkit): https://github.com/aimsresearchlab/ragscale
• Companion post: https://spanthi.com/blog/compression-is-a-coin-flip/

The file is byte-identical to data/interaction_matrix.csv.gz in the ragscale repository (version 0.5.1), which also bundles it:

from ragscale import load_interaction_matrix

matrix = load_interaction_matrix()

or, from this repository:

from datasets import load_dataset

matrix = load_dataset("vein05/ragscale-interaction-matrix", split="train")

Columns

Column	Type	Description
dataset	str	Source benchmark key: longmemeval, hotpotqa, musique, or nq
example_id	str	Item identifier within the stored benchmark slice
question	str	Question text
question_type	str	Source or analysis category when available
reference_answer	str	Stored reference answer
reader_model	str	Reader model identifier
method	str	Stored evidence-policy or diagnostic-condition label
correct	bool	Frozen binary outcome
judge_score	float	Stored raw judge score when available
generated_answer	str	Stored reader output
evidence_tokens	float	Stored evidence-token count when available
reader_raw_accuracy	float	Reader's binary score under the stored raw policy
transition_type	str	rescued, damaged, unchanged_correct, or unchanged_wrong for paired non-raw conditions

For one reader and item, rescued means wrong under the raw policy and correct under compression, and damaged means correct under the raw policy and wrong under compression. Raw-policy rows have an empty transition_type.

Scope

Dataset	Source items	Matrix rows	Paper role
LongMemEval	500	79,936	Semantic-judge replication
HotpotQA	500	45,500	Confirmatory RECOMP panel plus sensitivity policies
MuSiQue	500	20,000	Confirmatory shared-summary panel plus broader stored coverage
NQ-Open	324	31,428	Lexical boundary

The paper's NQ-Open analysis excludes one malformed source gold answer and reports 323 scored items. The 176,864 count covers this matrix only. It does not include the deterministic EM/F1 score sidecar or the later TriviaQA panel.

Correct use

The matrix includes auxiliary and sensitivity conditions beyond the fixed-artifact panels behind the paper's main claims. A shared method value does not prove that two readers received byte-identical compressed evidence. Before making a fixed-compressor claim, verify identical item footprints, identical candidate pools and compressed artifacts for every reader, the reader set, the metric version, and the resampling rule. The matrix alone cannot establish these; the paper's appendix and the ragscale repository describe the full artifacts.

License

The interaction matrix derives from separately cited public benchmarks (LongMemEval, HotpotQA, MuSiQue, NQ-Open) and stored model outputs. Benchmark content remains subject to its source license and terms. The ragscale code is released under Apache 2.0.

Citation

@misc{panthi2026ragcompression,
  title         = {Compression Is Not Evaluation-Neutral: Fixed {RAG} Compression Can Distort Reader Comparisons},
  author        = {Panthi, Sugam and Abdelfattah, Rabab},
  year          = {2026},
  eprint        = {2606.21807},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2606.21807},
  url           = {https://arxiv.org/abs/2606.21807}
}