ChihHsuan-Yang's picture
Correct Gemma paper-scope trajectory count (6,666 -> 5,930) and scope two claims
10336b1 verified
|
Raw History Blame Contribute Delete
12 kB

Schema — AAAI_Wrong-but-Useful private archive

Paper: Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages (arXiv:2608.14375).

Everything below is written for this archive. It supersedes docs/upstream_package/docs/schema.md, which describes the broader five-protocol scientific-agent-protocol-traces package.


0. The scope filter — read this before counting anything

The archive is a five-protocol superset. The paper studies one protocol. Every paper-specific number in the data card comes from this filter:

protocol_id == "dhd"
AND paper_analysis_eligible == True
AND actor_model_id IN ("gemma_4_31b", "gpt_oss_120b")

paper_analysis_eligible is the dataset's own discriminator for the strict complete-LOO analysis view. The actor_model_id clause is belt-and-braces: no gemma3_27b row is ever paper_analysis_eligible.

Two sibling filters exist and mean different things:

Column Meaning
paper_analysis_eligible strict complete-LOO analysis view — the paper's view
signal_analysis_eligible integrator received renderable hypothesis content
comparison_eligible the exact five-protocol benchmark matrix (appendix context)

1. Physical layout

Hive-style partitioning. Partition depth varies by table.

Path Partition keys
data/trajectories benchmark_id / actor_model_id / protocol_id
data/messages benchmark_id / actor_model_id / protocol_id
data/hypotheses benchmark_id / actor_model_id / protocol_id
data/interventions benchmark_id / actor_model_id / protocol_id
data/aggregate_metrics benchmark_id / actor_model_id / protocol_id
data/labels/outcomes benchmark_id / actor_model_id / protocol_id
data/labels/interventional_credit benchmark_id / actor_model_id
data/labels/matched_outcomes benchmark_id / actor_model_id
data/agents protocol_id / model_id
data/problems benchmark_id
raw/ benchmark_id / slice_id / model / protocol (DHD splits layer_a / loo)

2. Joins

problem_uid ──> trajectory_uid ──> (classic) event rows in data/messages
            └─> hypothesis_uid ──> intervention_uid ──> label_uid
  • problem_uid is the archive-internal stable key. source_problem_id is the benchmark-native identifier (kept in problems and labels/outcomes).
  • data/labels/interventional_credit.hypothesis_uid joins data/hypotheses.
  • data/labels/interventional_credit.intervention_uid joins data/interventions.
  • data/interventions.removed_hypothesis_uid names the hypothesis hidden in a leave-one-out replay; included_hypothesis_uids_json lists what remained.
  • raw_trace_pointer / raw_trace_path + raw_trace_row locate the originating record inside raw/**.jsonl.zst.

3. Tables

data/problems — 7,467 rows (5 files)

Benchmark problem records including question text and gold answers. Partitioned by benchmark_id only. 6,666 of these are in paper scope; the 801-row excess is LAB-Bench's text_no_tool slice (see §5).

data/trajectories — 82,224 rows (paper scope: 12,596)

One row per end-to-end solve attempt. Key columns: trajectory_uid, problem_uid, source_problem_id, benchmark_id, slice_id, protocol_id, actor_model_id, evaluator_model_id, run_id, comparability_group_id, compute_site, inference_backend, source_commit, prompt_version, final_correct, first_pass_correct, final_answer, total_tokens, prompt_tokens, completion_tokens, model_calls, evaluator_calls, wall_time_seconds, answer_parse_status, event_count, message_count, raw_trace_path, raw_trace_row, source_schema, provenance_status, release_status, trace_scope, comparison_eligible, paper_analysis_eligible.

Token/time columns are NULL on every DHD row (0 of 13,332 populated). Cost figures exist only for the classic protocols. Do not compute a DHD cost from this table.

data/messages — 1,889,265 rows — ZERO DHD ROWS

Ordered classic-protocol trace events. protocol_id takes exactly four values: baseline_llm, broadcast, per, single_agent. Verified by a distinct-value scan of the whole column.

This table is not the paper's "messages." See docs/label_card.md §1.

data/hypotheses — 66,660 rows (paper scope: 62,445)

This is the paper's "message" object. One row per generated hypothesis; 6,666 problems x 5 hypotheses x 2 actor models = 66,660. Key columns: hypothesis_uid, problem_uid, benchmark_id, slice_id, protocol_id, actor_model_id, evaluator_model_id, run_id, hypothesis_index, agent_uid, role_name, role_description, reasoning_lens, proposed_answer, proposed_answer_extracted, reasoning_direction, key_idea, critical_assumption, what_to_check, confidence, parse_complete, proposal_correctness, proposal_correctness_class, proposal_answer_status, operational_proposal_success, proposal_judge_rationale, hypothesis_text_hash, judge_model, prompt_version, source_commit, compute_site, signal_analysis_eligible, comparison_eligible, paper_analysis_eligible.

The six integrator-visible content fields are proposed_answer, reasoning_direction, key_idea, critical_assumption, what_to_check, confidence. Signal eligibility requires at least one to be renderable.

data/interventions — 333,300 rows (paper scope: 312,781)

One row per requested replay.

intervention_type rows meaning
k0_no_hypotheses 13,332 integrator with no message (the control)
individual_hypothesis 66,660 integrator with exactly one message
full_committee 66,660 integrator with all five
leave_one_out 186,648 integrator with one of five hidden

measurement_status is observed (319,764) or missing (13,536). Missing is never coerced to neutral — it carries a categorical missing_reason and produces no interventional_credit row.

There are exactly four intervention types. The component-masking conditions (answer_masked, reasoning_masked, neutral_replacement, byte_identical_repeat) are not present — see docs/datasheet.md §7.

data/labels/interventional_credit — 241,256 rows (paper scope: 237,205)

The paper's core measurement. signed_effect = outcome_with - outcome_without in {-1, 0, +1}.

credit_scope paper-scope rows credit method
committee_hypothesis 174,760 leave_one_out_marginal_effect
individual_hypothesis 62,445 single_hypothesis_vs_k0

Key columns add credit_scope, credit_method, committee_size, proposal_correctness, proposal_correctness_class, outcome_with, outcome_without, signed_effect, effect_class, wrong_but_useful, correct_and_useful, harmful.

data/labels/outcomes — 82,224 rows (paper scope: 12,596)

Scored final outcome per trajectory; same key set as trajectories plus evaluation_mode, evaluation_source, label_version.

data/labels/matched_outcomes — 17,223 rows — cross-protocol by construction

One row per problem x actor model, with a column per protocol (baseline_correct, single_correct, per_correct, broadcast_correct, dhd_correct plus matching *_tokens), oracle_cheapest_successful, oracle_token_cost, any_protocol_solved, five_protocol_traces_present, five_protocols_complete. It has no paper_analysis_eligible column and no DHD-only version: filtering it to "paper scope" is not meaningful.

data/aggregate_metrics — 62 rows (DHD: 10)

Per-cell rollup: n, n_labeled, n_missing_outcome, n_correct, final_accuracy, avg_total_tokens, avg_model_calls, avg_wall_time_seconds, release_status. Cost columns are populated on only 16 of 62 rows.

data/agents — 30,976 rows (DHD: 30,768)

Agent/role records, partitioned by protocol_id / model_id.

registry/ — 9 parquet files + 3 metadata files

benchmarks (5), models (3), protocols (5), licenses (5), label_definitions (6), runs (62), coverage (62), dhd_measurement_coverage (10), problem_crosswalk (17,371), quarantine (0 rows — nothing quarantined), plus release_manifest.json, checksums.sha256, schema.json.

registry/dhd_measurement_coverage.parquet independently reproduces the paper's tab:app_analysis_scope: loo_complete_problems = 3,724 / 500 / 577 / 482 / 647 (total 5,930) and loo_eligible_problems total 5,980.

raw/ — 72 zstd JSONL shards, 94,870 rows (82,224 referenced)

Schema dhd-sanitized-canonical-v1. Top-level keys: actor_model_id, benchmark_id, comparison_eligible, complete_loo, evaluator_model_id, outcome, problem_uid, protocol_id, provenance, run_id, schema_version, slice_id, source_problem_hash, source_problem_id, trace, trajectory_uid. provenance retains compute_site, inference_backend, prompt_version, judge_model — no credentials, no filesystem paths.

viewer_samples/ — 30 JSONL files

Six protocol/model views x five benchmarks, for quick browsing.

paper_derived/ — the arXiv ancillary reproducibility artifact (85 entries)

The authors' public-safe derived layer, verbatim. Contains data/derived/*.csv|json, figures/*_source.csv (the figure inputs), configs/paper/dhd_config*.yaml (prompts + decoding), src/dhd/, scripts/, tests/. All 69 hashes in its own SHA256SUMS re-verify in place.


4. Models — actors vs evaluators

model_id display role in this paper DHD trajectories
gpt_oss_120b gpt-oss-120b ACTOR (OSS) + evaluator 6,666
gemma_4_31b gemma-4-31B-it ACTOR (Gemma) 5,930
gemma3_27b gemma-3-27b-it NOT IN THIS PAPER'S DHD ARM (registry role is actor; 3,332 classic-protocol trajectories; evaluator_model_id 0 times) 0
Meta-Llama-3.1-70B-Instruct Llama-3.1-70B EVALUATOR ONLY not in this archive

Gemma is 5,930 not 6,666 under the paper-scope filter: 736 Gemma DHD trajectories are paper_analysis_eligible == False (incomplete replay records); OSS has 0. See registry/dhd_measurement_coverage.parquet column paper_analysis_problems.

evaluator_model_id is gpt_oss_120b on every row of every table in this archive. The three-way evaluator-agreement analysis is therefore not reproducible from these tables; its published result ships as paper_derived/figures/evaluator_agreement_source.csv.


5. The 7,467 -> 6,666 reconciliation

Benchmark problems table in paper scope excess
omnimath2 4,181 4,181 0
jeebench 515 515 0
scibench 580 580 0
mascqa 649 649 0
labbench 1,542 741 801
Total 7,467 6,666 801

The entire excess is LAB-Bench. llm_strict (CloningScenarios 33 + ProtocolQA 108 + SeqQA 600 = 741) is the paper's headline slice; the broader 1,542-row text_no_tool slice is a separate closed-book stress test. Split by available_slices_json: ["llm_strict","text_no_tool"] -> 741 rows, all in scope; ["text_no_tool"] -> 801 rows, none in scope.


6. Worked example

import pyarrow.dataset as ds, pyarrow.compute as pc

ROOT = "."   # local snapshot of this private repo
PAPER = ((pc.field("protocol_id") == "dhd")
         & (pc.field("paper_analysis_eligible") == True)
         & (pc.field("actor_model_id").isin(["gemma_4_31b", "gpt_oss_120b"])))

credit = ds.dataset(f"{ROOT}/data/labels/interventional_credit",
                    format="parquet", partitioning="hive")

# the paper's headline LOO event counts: 91,740 OSS / 83,020 Gemma
loo = credit.to_table(filter=PAPER & (pc.field("credit_scope") == "committee_hypothesis"),
                      columns=["actor_model_id", "benchmark_id", "signed_effect",
                               "proposal_correctness_class", "wrong_but_useful"])