tada-bench / README.md
jxtse's picture
README: gated access note
d820cd7 verified
|
Raw History Blame Contribute Delete
3.11 kB
metadata
license: other
pretty_name: TADA-Bench (pilot)
tags:
  - benchmark
  - agents
  - text-analysis
  - harbor

TADA-Bench pilot tasks

Nine professional text-analysis tasks produced by the TADA task-generation pipeline (https://github.com/jxtse/tada, benchmark/taskgen/). Each task combines one to three source datasets into a realistic work request whose answers decode one-to-one back to the original dataset labels, so every task is scored deterministically against original ground truth.

Each directory under tasks/ is a standard Harbor task:

task.toml  instruction.md  environment/  tests/  solution/

Run one task with Harbor's reference solution (no model):

hf download jxtse/tada-bench --repo-type dataset --local-dir tada-bench
harbor run -p tada-bench/tasks/<task-id> -a oracle

Tasks run offline (network_mode = no-network); the agent writes /output/review.xlsx.

Tasks

tasks.json lists, for every task: source datasets, difficulty stratum, the four pilot trial rewards (DeepSeek-Flash x2, Claude Opus x2), the oracle reward, and a file-tree SHA-256.

Task Sources Difficulty
maud-pilot MAUD easy
bird-infiagent-oolongsynth BIRD, InfiAgent-DABench, Oolong-synth medium
dabstep-semtpch-spider2 DABstep, SemTPCH, Spider2 hard
prec-loft-lrobench PREC (SCDB/CAP), LOFT, LRO-Bench hard
infiagent-kramabench-palimpzest InfiAgent-DABench, KramaBench, Palimpzest easy
biodex-udabench BioDEX, UDA-Bench hard
blade-udabench BLADE, UDA-Bench hard
b-dabstep-tablebench Evidence Inference, DABstep, TableBench medium
cuad-maud-lrobench CUAD, MAUD, LRO-Bench easy

Difficulty is a stratum, not an admission threshold: easy = at least 3 of 4 trials reach reward 0.8, medium = 2, hard = fewer than 2.

Validation

  • Six offline hard gates per task: oracle invertibility, trivial-answer baselines, determinism, leakage, keyword baseline, output contract.
  • Independent model review (instruction sufficiency, scoring fairness, realism, text-analysis difficulty).
  • Four agent trials in Harbor; cells that every trial missed were blind-read twice, and confirmed source-label conflicts were excluded (never relabelled) before a single rebuild.
  • Harbor oracle agent: 9/9 tasks score 1.0.

Licensing and access

Access is gated: request access on this page; the maintainer approves requests manually. Task packages contain text, tables and labels from the source datasets, which keep their own licences: CUAD and MAUD (CC BY 4.0), SCDB/PREC (CC BY-NC 3.0 US), BIRD (CC BY-SA 4.0), BioDEX (per-article full-text licences, including NC/ND), DABstep (CC BY 4.0), Spider2/UDA-Bench/Oolong/Palimpzest/Evidence Inference (MIT), LOFT/SemTPCH/TableBench/BLADE/ InfiAgent-DABench (Apache-2.0, with upstream data under their own terms); KramaBench and LRO-Bench publish no licence. Do not redistribute until licences are cleared per dataset.

Known limitation: public datasets may have been seen by models during pre-training.