Benchmark grades — scores only, no text, no PII
The judged panel behind the harness-lift result. Each prompt is answered by the same model bare and harnessed, then scored 0-100 by a three-judge panel (each judge from a different model family, never grading its own) across five reasoned safety dimensions. These files are the source of every headline number.
Panel grades
Harness benchmark grades
One row per (model, arm, prompt, judge): the overall 0-100 score plus the five A-E component scores. The primary evidence file.
85,417grade rows · 7,973 prompts · 8 models
model, arm, prompt_id, judge, score_0_100, A, B, C, D, E
Per-dimension grades
Per-dimension grades (exhaustive)
Higher resolution: one reasoned judge call per rubric dimension (A indicator, B legal, C refusal, D resources, E privacy), re-versioned as the exhaustive sweep grows toward the full registry.
growingone judge call per dimension
model, prompt_id, arm, judge, score_0_100, comp_A..comp_E
Cross-model board
Cross-model leaderboard
A citable flat table of every model's baseline, harnessed mean, raw lift, ceiling-adjusted normalized gain, and win rate. The two rankings genuinely reorder.
8models ranked
model, baseline, harnessed, raw_lift, norm_gain, win_rate, n_pairs
Controls
Methodology controls
The placebo panel, negative control, and applicability audit - including the honest, inconclusive parts. The evidence that the lift is not a scoring artifact.
3control studies
study, arm, prompt_id, metric, value
Raw prompts and responses — synthetic, scrubbed, text-bearing
Where the grades are scores only, these datasets carry the actual language: the adversarial prompts and the model's answers. Prompts are synthetic composite scenarios (no real individual), kernel metadata is scrubbed, and rows tripping a conservative PII scan are dropped. Built for natural-language processing, sentiment, and keyword work.
Prompt + responses
Prompt and response showcase
The raw adversarial prompt plus the model's baseline, harness-core, and harness-full answers, side by side. Read what the harness actually changes, or run your own NLP on it.
1,087prompts × 3 responses
prompt_id, category, corridor, difficulty, prompt_text, baseline_response, harness_core_response, harness_full_response
Chain of thought
Chain-of-thought reasoning
Multi-perspective reasoning chains grounded in the ILO forced-labour indicators: 101 perspectives across 9 role categories, each a ~102-step chain, with a clean held-out split. Silver labels for reasoning fine-tuning.
2,020chains · 1,740 train / 280 held out
messages, perspective, category, situation, ilo_indicator, reach, direction, step_count, lineage_id, split, ...
Training corpora — manifest-bound, held-out splits
The distilled fine-tuning data: measured-response supervised pairs, multi-perspective coverage across roles, time, evidence, and jurisdiction, and the proof corpus attached to the fine-tuning kernel. Data cards, schema and loading guides, Croissant metadata, checksums, and publication boundaries are packaged beside the rows.
Supervised responses
Measured response training corpus
Graded worst-to-best response examples distilled into supervised fine-tuning and preference rows, with a quarantine split for dirty rows.
Multi-perspective
Multiperspective fine-tuning corpus
Coverage across declared synthetic axes - roles, time frames, evidence states, and jurisdictions - so a fine-tune does not collapse to one point of view.
Proof corpus
Proof fine-tuning data
The compact corpus attached to the fine-tuning and evaluation kernel for the on-Kaggle Low-Rank Adaptation smoke run.
Reproduce it yourself — the analysis notebooks
Every dataset above ships with a public, central-processing-unit-safe notebook that loads it live and recomputes the figures - distributions, the headline lift, the cross-model board, per-dimension gains, judge agreement, and the honest boundary. No hidden state; attach the dataset and re-run any cell.
Start here
Benchmark index
The front door: the headline result, the cross-model board, and a guided tour of the whole collection.
Flagship
Does a safety harness help?
The grandmaster analysis: score-shift, the per-prompt lift, the cross-model board, a per-dimension radar and heatmap, judge agreement, and difficulty and category breakdowns.
NLP
Prompt and response NLP explorer
Raw prompt and response row by row, sentiment shift, distinctive vocabulary (baseline vs harnessed), citation and refusal detection, readability, and wordclouds.
More
Explorers, deep-dives, and data cards
Per-dimension explorer, cross-model leaderboard deep-dive, chain-of-thought explorer and analysis, and schema data cards for every release.
Load any of it in three lines
Attach a dataset to a Kaggle notebook (or download the CSV) and read it with pandas. Grades pair per prompt; text datasets read straight into a data frame.
# in a Kaggle notebook: Add data -> search "duecare" -> attach
import pandas as pd, glob
grades = pd.read_csv(glob.glob("/kaggle/input/**/panel_grades.csv", recursive=True)[0])
# mean the judge panel per (prompt, arm), then pair harnessed vs baseline
piv = grades.groupby(["prompt_id", "arm"])["score_0_100"].mean().unstack()
lift = (piv["harness_core"] - piv["baseline"]).dropna()
print(f"mean lift +{lift.mean():.1f} over {len(lift):,} paired prompts, "
f"{100 * (lift > 0).mean():.1f}% improved")
Grades and per-dimension scores carry no response text and no personal data. The prompt-and-response and chain-of-thought datasets are synthetic composite scenarios with kernel metadata scrubbed and a conservative personal-data scan applied. Everything here is open-licensed (grades and scores under a Creative Commons license; corpora carry their own license files). Judges are language models and prompts are synthetic - this is benchmark evidence about response quality, not ground truth about any person. Source and build scripts: github.com/TaylorAmarelTech/gemma4_comp.