VERSIONS.md

Dataset versions and what they back

Ties a released dataset version to the experiments it produced and the paper tables those experiments back, so a reader can go from a number in the PDF to the exact prompts a model saw.

anchorbench-v2.0-core

Git tag	colm2026-camera-ready
Suites	external, history, icl, rag, tool (360 items each) + external_uncertain
Released files	datasets/hf_release/*.jsonl (14,400 rows)
Hub	Yiderigun/LLM_anchoring (currently private)
Checksums	datasets/anchorbench_core_checksums.sha256
Raw generations	tarballs on Google Drive
HF revision	85b6d4a26092868281f45b745fd1f5ccf02dd91b (2026-08-05)

Generated with, for every suite:

python -m anchorbench.data.generate --suite <suite> --size core --seed 42

external_uncertain is derived rather than generated: the runner re-renders each external item at k = 1, 2, 3 visible ratings (anchorbench.runners.rebuttal_uncertain.build_uncertain_promptviews).

Which results came from it

Results directory	Models	Backs
results/full_benchmark/	10 open-weight	Table 1, Figs 3-4, most appendix tables
results/api_benchmark/	4 frontier API	Table 1, Figs 3-4
results/rebuttal/uncertain/	4 open-weight	Table 2 (main paper)
results/rebuttal/**	varies	the 13 \input-ed appendix tables

Provenance chain

datasets/anchorbench_<suite>_core/   generated, seed 42, committed
  -> results/<experiment>/           raw generations, Drive + CHECKSUMS.sha256
    -> outputs/tables/*.tex          regenerated by `anchorbench tables`
      -> COLM_camera_ready/          the PDF, at tag colm2026-camera-ready

Every link is checked by a test: tests/test_dataset_regeneration.py for the first, tests/test_golden_artifacts.py for the third, and anchorbench verify for the last.

Reproducibility notes carried by this version

• History anchor_value is null by design. History's anchor is the model's own Stage-1 answer, so it is a property of the (item, model) run, not of the item. The same item records 80 for Qwen-7B and 81 for Llama-8B. Read the realised value from the released results.
• split is test everywhere. AnchorBench is eval-only; there is no train or validation portion. The internal itemspecs carry tags.split = "core", which is a size label and is deliberately not propagated, since the Hub would render it as a data split.
• Three prompts repeat inside uncertain_p1_control (6 of 360 items, 1.7%) with different gold answers. This is the design, not a collision: at k=1 the prompt shows a single visible rating while gold remains the mean of all five, so two items sharing a first rating produce identical prompts with different answers. That irreducible uncertainty is what the k=1 condition exists to create.
• vLLM is not bitwise reproducible at temperature 0. Re-running a cell can move a metric by a few hundredths; see results/rebuttal/cot_replication/README.md.

Superseded Hub revisions

fbbadc1ed373 (2026-07-10) — the first upload. Its prompts were correct and byte-identical to the evaluated ones, but it carried only five fields (item_id, condition, prompt_text, y_star, anchor_value), which is not enough to reproduce the paper's own breakdowns: Figure 3 needs offset and tab:difficulty needs difficulty. It also published anchorbench_tool_read_core/, a suite removed in v2.0 that appears nowhere in the paper, and omitted external_uncertain, which backs Table 2 in the main paper. Replaced in one commit by 548dff842706, then 85b6d4a26092 corrected the homepage URL. The old files remain in the Hub's git history.