tomato-egg-bench / README.md
fancyboi999's picture
Release TomatoEggBench-120 dataset and community split
c6d566d verified
|
Raw History Blame Contribute Delete
2.88 kB
metadata
license: apache-2.0
task_categories:
  - text-classification
tags:
  - agent-evaluation
  - decision-model
  - system-one
  - jev
  - tomato-egg-problem
  - benchmark
size_categories:
  - n<1K

TomatoEggBench: Evaluating Final-State Discipline in AI Agent Deliverables

"Ship the result, not the conversation."

Hugging Face Models GitHub


Dataset Summary

TomatoEggBench is a 120-item public benchmark designed to test whether AI agents deliver clean, invariant final states or leak conversational scaffolding, discarded alternatives, and negative echoes into user deliverables.

The dataset is named after the "Tomato-Egg Problem" highlighted by UC Berkeley researcher Shangyin Tan:

"I asked an AI agent to write a recipe for tomato scrambled eggs. It proudly delivered: 'Here is your tomato scrambled eggs (note: this dish contains no pork, no beef, no chicken, and no fish).' Why are AI agents so obsessed with telling us what they didn't do? Ship the result, not the conversation."

Data Composition

The benchmark covers 5 major engineering and enterprise deliverable domains:

  1. Code Repositories: Commit messages, PR descriptions, code comments, and docstrings.
  2. Spreadsheets & Data: Financial models, CSV/Excel summaries, accounting exports.
  3. Formal Documents & PDFs: Technical specifications, executive summaries, compliance filings.
  4. Form Automation & RPA: Robotic process automation input data, submission manifests.
  5. Executive Slide Decks: Client pitches, strategy decks, board presentations.
Split Description Total Items Pass Rate Revise Rate
test.jsonl Standard Held-Out Benchmark 120 50.0% (60) 50.0% (60)
community_40.jsonl Real Community Bug Reports 40 37.5% (15) 62.5% (25)

Key Findings on TomatoEggBench-120

Model Architecture Accuracy Violations Caught (Recall) Latency
Ockev-3B Qwen2.5-3B + Pointer Head 95.8% 98.3% 86 ms
TypeSafe Jev 1.13.0 Causal MoE + RLCD 93.3% 96.7% 250 ms
Ockev-1.5B Qwen2.5-1.5B + Pointer Head 92.5% 95.0% 35 ms
SemIf Qwen3.5-4B Qwen3.5-4B (Logits Readout) 74.7% 68.3% 850 ms
ModernBERT-151M Sequence Classification Head 25.0% 35.0% 24 ms

Usage

from datasets import load_dataset

dataset = load_dataset("fancyboi999/tomato-egg-bench")
print(dataset["test"][0])

Citation & Code

Official repo: https://github.com/fancyboi999/ockev