anonymity11's picture
Add eval-200 README
76b3c63 verified
|
raw
history blame
2.81 kB

FormulaCascade Benchmark — Evaluation Subset (eval-200, v2)

This is a 200-task stratified evaluation subset of the full FormulaCascade Benchmark (11,171 tasks).

Sampling Design (v2)

  • 200 tasks total: 50 per tier × 4 tiers
  • Balanced: 25 single-sheet + 25 cross-sheet per tier
  • Quality-filtered: all tasks are verdict=RELEASE, total QC score ≥ 26, D3 ≥ 3, D4 ≥ 3
  • Diverse: maximizes domain and formula-family coverage within each tier
  • Reproducible: sampled with seed=42 using resample_eval200_v2.py

v2 Improvements over v1

  • All 200 tasks are verdict=RELEASE (v1 had 31 HOLD tasks in T4_STANDARD)
  • Minimum QC score raised to 26/30 (v1 allowed 25)
  • T2_HARD: removed extreme outliers with total_masked > 2000
  • T4_STANDARD: depth-stratified (depth=2: 70%, depth=3–4: 20%, depth=5+: 10%) for better differentiation from T3_MEDIUM
  • T1_FRONTIER depth=2 tasks require total_masked ≥ 50 to ensure genuine difficulty

Tier Distribution

Tier Count Single Cross Depth Range Avg Depth QC Score (avg)
T1_FRONTIER 50 25 25 2–45 10.5 27.6
T2_HARD 50 25 25 3–4 3.4 27.8
T3_MEDIUM 50 25 25 2 2.0 27.8
T4_STANDARD 50 25 25 2–8 2.7 26.0

Domain & Formula Coverage

Domain Count Formula Family Count
general 35 arithmetic 68
finance 34 conditional 42
sales 31 aggregation 39
data_analysis 28 lookup 33
education 19 ranking 6
operations 19
hr 17
engineering 17

Directory Structure

FormulaCascade_Benchmark_eval_200/
├── README.md
├── manifest_eval200.jsonl        # 200-task metadata (v2)
├── fcb_id_mapping.json           # FCB-ID ↔ task_id mapping (200 tasks)
├── eval200_summary_v2.json       # sampling statistics summary
├── T1_FRONTIER/
│   ├── release/FCB-NNNNN/        # task_description.json + task-NNNNN.xlsx
│   └── internal/FCB-NNNNN/       # full provenance + ground truth
├── T2_HARD/
├── T3_MEDIUM/
└── T4_STANDARD/

Usage

For evaluation, use release/FCB-NNNNN/task_description.json (prompt) and release/FCB-NNNNN/task-NNNNN.xlsx (masked workbook) as solver input.

Ground truth is in internal/FCB-NNNNN/ground_truth_target.json (layer-1 target cells).

See the full benchmark README for evaluation protocol details.

Tier Counts

  • T1_FRONTIER: 50 tasks (all RELEASE, depth 2–45, avg depth 10.5)
  • T2_HARD: 50 tasks (all RELEASE, depth 3–4)
  • T3_MEDIUM: 50 tasks (all RELEASE, depth 2)
  • T4_STANDARD: 50 tasks (all RELEASE, depth 2–8)