Datasets:

Amine-CV
/

whitecircle_killbench_analysis

pretty_name: White Circle KillBench — Reasoning-Trace Analysis
language:
  - en
tags:
  - llm
  - ai-safety
  - bias
  - fairness
  - llm-evaluation
  - reasoning
  - chain-of-thought
  - interpretability
  - killbench
  - thinkprobe
size_categories:
  - 100K<n<1M
task_categories:
  - text-classification
configs:
  - config_name: features
    data_files: data/features_labelable.parquet
    default: true
  - config_name: traces
    data_files: data/pop_labelable.parquet
  - config_name: labels
    data_files: data/labels.parquet
  - config_name: choices
    data_files: data/choices_all.parquet
  - config_name: refusal_traces
    data_files: data/pop_refusal.parquet
  - config_name: refusal_features
    data_files: data/features_refusal.parquet
  - config_name: bias_groups
    data_files: data/bias_groups.parquet
  - config_name: metric_sweep
    data_files: data/comparisons_metric_results.parquet
  - config_name: profile_sweep
    data_files: data/comparisons_profile_results.parquet
  - config_name: decision_stability
    data_files: data/behaviour_stability_full.parquet

White Circle KillBench — Reasoning-Trace Analysis

Derived data from an independent analysis of whitecircle/killbench, White Circle's benchmark of demographic bias in life-or-death decisions. Every reasoning trace from that dataset was profiled with ThinkProb, which turns a trace into a thought graph and computes 22 structural metrics without generating any text.

📄 Code, reports and figures: github.com/kmamine/whitecircle_killbench_anlaysis
🗃️ Source dataset: whitecircle/killbench

Attribution: "Dataset provided by WhiteCircle.ai under the White Circle Responsible Use License."

No KillBench text is included. Each row keeps KillBench's own row_id, so it can be joined one-to-one with the source dataset to get the prompts, responses and reasoning traces (see Joining back to KillBench).

Headline results


84.8%	gemini-3.1-pro picks whoever is listed first (uniform choice would be 25%)
29.6%	the best model picks the same person when a dilemma is repeated with the order shuffled (chance: 6.25%)
72%	the person with no phone is picked to die, with no extra hedging in the reasoning
\|β\| ≤ 0.036	trace structure vs. bias, after controlling for length and lexical features: a controlled null

Configs

Config	File	Rows	Contents
`features` (default)	`data/features_labelable.parquet`	69,867	22 ThinkProb metrics, graph size, and lexical baseline features for each trace
`traces`	`data/pop_labelable.parquet`	69,867	Metadata and choice for each trace in the analysed population (text removed)
`labels`	`data/labels.parquet`	69,867	Bias-alignment labels for each trace
`choices`	`data/choices_all.parquet`	552,973	Every structured choice across all 15 models (used for position bias and repeatability)
`refusal_traces` / `refusal_features`	`data/pop_refusal.parquet` / `data/features_refusal.parquet`	13,351	Long refusal traces and their ThinkProb features
`bias_groups`	`data/bias_groups.parquet`	1,365	Bias size for each model × dimension × scenario (G-test, Cramér's V)
`metric_sweep`	`data/comparisons_metric_results.parquet`	9,420	Every bias↔metric test (β, partial r, BH q-values)
`profile_sweep`	`data/comparisons_profile_results.parquet`	1,192	Every bias↔profile-family analysis
`decision_stability`	`data/behaviour_stability_full.parquet`	13,120	Same person / same slot across the 3 shuffled repeats of each dilemma

The rest of data/ holds the intermediate tables behind each report: analysis, behaviour, chain, profiles, kill rates and findings. tables/ holds the CSV and JSON summaries that the GitHub reports quote.

Analysed population

The traces come from KillBench rows with structured output, ≥ 2,000 characters of reasoning, a single varied dimension, no refusal and a parsed choice: 69,867 traces in total. Only three models produce reasoning of that length: moonshotai/kimi-k2.5, qwen/qwen3-235b-a22b and google/gemini-3.1-pro-preview.

Columns

Trace metadata (traces, choices): row_id, model_id, varied_param (the bias dimension), scenario_id, scenario_domain, language, group_id, setup_id, roll_idx (which of the 3 shuffled orders), rlen (reasoning length in characters), chosen_value, available_values (the 4 people offered).

ThinkProb metrics (features), in 5 families plus 2 length controls:

Family	Metrics
Breadth	`branching_factor`, `unique_perspective_count`, `domain_spread`, `first_idea_diversity`
Depth	`max_elaboration_chain`, `mean_branch_depth`, `specificity_gradient`, `reasoning_density`
Structure	`exploration_exploitation_ratio`, `backtracking_rate`, `cross_branch_connectivity`, `convergence_index`, `graph_density`, `revision_depth`
Metacognitive	`self_reflection_rate`, `critique_to_hypothesis_ratio`, `hedging_density`, `perspective_taking`
Efficiency	`token_per_idea`, `redundancy_ratio`
Length controls	`avg_tokens`, `avg_tus`

Also included: graph size (n_tus, n_edges) and the lexical baseline:

length: lex_char_len, lex_word_count, lex_sent_count, lex_words_per_sent;
hedging and boosters: lex_hedge_count, lex_hedge_rate, lex_booster_rate;
other word rates: lex_negation_rate, lex_connective_rate, lex_first_person_rate;
lex_question_count.

Labels (labels): for each trace, how strongly the chosen value matches the model's overall preference. This is computed at two levels — model × dimension (*_md) and model × dimension × scenario (*_mds). Each level has three columns: how many times the value was on offer (avail_*), how much it is preferred (align_*), and whether it passed the minimum sample-size check (valid_*). with_bias says whether the choice goes the same way as the bias.

Joining back to KillBench

row_id (for example run_001:000000) is KillBench's own primary key. It is unique across all 1,368,936 source rows, and every row_id in this dataset exists in the source.

import pandas as pd
from datasets import load_dataset

kb = load_dataset("whitecircle/killbench", split="train",
                  columns=["row_id", "reasoning_text", "response_text"]).to_pandas()
traces = load_dataset("Amine-CV/whitecircle_killbench_analysis", "traces", split="train").to_pandas()
feats  = load_dataset("Amine-CV/whitecircle_killbench_analysis", "features", split="train").to_pandas()

full = traces.merge(feats, on="row_id").merge(kb, on="row_id", how="left")   # 69,867 rows

Limitations

Three models only. Only kimi-k2.5, qwen3-235b and gemini-3.1-pro produce reasoning long enough to profile.
Choices are parsed by a judge, and bias labels are built at the group level rather than per row.
The trace and the choice come from the same generation, so a link between them is a correlation, not proof that one causes the other.
ThinkProb was built for open-ended reasoning; KillBench traces are shorter and decision-style.
The reasoning may not reflect what the model actually computes.

Cite this

Dataset (derived data on HuggingFace):

@misc{kerkouri2026killbench_data,
  author       = {Kerkouri, Mohamed Amine},
  title        = {White Circle KillBench: A Reasoning-Trace Analysis (Derived Data)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Amine-CV/whitecircle_killbench_analysis}},
  note         = {ThinkProb structural features, labels and choices for KillBench reasoning traces, keyed by KillBench \texttt{row\_id}}
}

Code (analysis pipeline on GitHub):

@software{kerkouri2026killbench_code,
  author = {Kerkouri, Mohamed Amine},
  title  = {White Circle KillBench: A Reasoning-Trace Analysis (Code)},
  year   = {2026},
  url    = {https://github.com/kmamine/whitecircle_killbench_anlaysis}
}

Source benchmark (White Circle's KillBench — please cite it too):

@misc{whitecircle2026killbench,
  author       = {{White Circle}},
  title        = {{KillBench}: Discovering Hidden Biases of {LLMs}},
  year         = {2026},
  month        = apr,
  publisher    = {Hugging Face},
  url          = {https://whitecircle.ai/killbench},
  howpublished = {\url{https://huggingface.co/datasets/whitecircle/killbench}}
}