luoojason/jsaj-eval-bundle

JSAJ Eval Bundle

Pre-materialized cells for the JSAJ team's context degradation evaluation on the MMLongBench-Doc derived dataset. Each cell folder contains the exact PDFs an evaluation must run against.

Structure

q0/
  control_k0/        source.pdf + question.json
  hard_negative_k2/  source.pdf + hn_1.pdf + hn_2.pdf + question.json
  hard_negative_k4/  source.pdf + hn_1.pdf .. hn_4.pdf + question.json
  random_k2/         source.pdf + random_1.pdf + random_2.pdf + question.json
  random_k4/         source.pdf + random_1.pdf .. random_4.pdf + question.json
q1/
  ...
manifest.json        cell_id -> folder path lookup

286 questions, 5 conditions each = 1,430 cells. Source PDF is always position 0.

Cells

• control_k0 — source document only
• hard_negative_k2 / k4 — source + 2 or 4 topical HN PDFs (curated per question)
• random_k2 / k4 — source + 2 or 4 HN PDFs sampled from other questions (one HN per other question; never from the source question's own HN pool)

Distractor selection is deterministic, seeded at 20260523. Cells are byte-identical for every model that consumes this bundle.

Per-cell metadata

question.json in each cell folder contains:

{
  "cell_id": "q0_random_k4",
  "doc_id": "PH_2016.06.08_Economy-Final.pdf",
  "question": "...",
  "answer": "...",
  "answer_format": "Str",
  "condition": "random",
  "k": 4,
  "n_pages": 165,
  "est_tokens": 66000,
  "bundle_filenames": ["source.pdf", "random_1.pdf", ...],
  "original_filenames": ["PH_2016...pdf", "web_e6e2...pdf", ...]
}

How to use

import json
from pathlib import Path

# load a single cell
cell = json.load(open("q0/random_k4/question.json"))
pdfs = [Path(f"q0/random_k4/{name}") for name in cell["bundle_filenames"]]
# send pdfs to your model with cell["question"]; score against cell["answer"]

For batch evaluation, iterate over manifest.json to get all cell folders.

Source dataset

Derived from MMLongBench-Doc with a 286-row text-safe filter. Source + HN PDFs originally curated at luoojason/mmlongbench-text-only.

Citation

Part of the JSAJ team's Algoverse March 2026 cohort research on long-context degradation. Code: SaibililaA/JSAJ.