nishan-chatterjee/llm-bias-detection

LLM Bias Detection Evaluation Traces

Evaluation data accompanying Navigating the Digital Spectrum: Assessing Political Bias, Moral Values, and Toxicity in LLMs.

The Dataset contains the model outputs used for Political Compass multiple choice and chat experiments, IBM topic sentiment classification, and identity-targeted hate-speech detection. Reproduction and analysis code is in the LLM Bias Detection GitHub repository.

Models and evaluation conditions

The primary model set is:

• instruction-tuned Gemma 3 1B, 4B, 12B and 27B from the Gemma 3 release;
• Qwen3 4B, 8B, 14B and 32B from the Qwen3 collection, including matched thinking and non-thinking chat protocols.

The optional ablation configuration uses YanLabs/gemma-3-27b-it-abliterated-normpreserve-v1. Pinned repository revisions and decoding parameters are provided in metadata/model_config.json. No model weights are included.

Dataset configurations

political_compass_mcq

Twenty-four Parquet files: eight primary models evaluated in bf16, 8-bit and 4-bit conditions. Each file has 1,800 sampled prompt configurations covering six assigned-persona conditions. The experiment spans fourteen languages.

Rows contain experimental factors and the four candidate probabilities for each of 62 propositions. Historical files did not record complete vocabulary logits. The learned Political Compass scorer is not included; released configuration-level coordinates and reconstruction code are available in the GitHub repository.

political_compass_chat

Twelve Parquet files: four Gemma chat variants and four Qwen checkpoints in think/no-think modes. Each variant has 111,600 question rows: 1,800 prompt configurations × 62 propositions.

Rows include the prompt, visible Stage-1 response, token count, finish reason, candidate log-probabilities and candidate probabilities. Stage 2 scores the answer candidates and does not produce a second rationale. Qwen thinking traces contain the visible model output returned by the experiment; they should not be treated as privileged hidden reasoning.

political_compass_chat_ablation

A separate matched Gemma 3 27B abliterated-checkpoint diagnostic. It is not part of the eight-model primary comparison.

ibm_sentiment

Eight Parquet files with 54,000 rows each: 1,800 prompt configurations × 30 unique IBM topic-sentiment items. Rows include the topic, target phrase, gold positive/negative label, model prediction, correctness, candidate scores, assigned persona and prompt factors. This MCQ task has no generated rationale.

The release analysis uses gold labels and experiment fields only. It does not include the discarded LLM-authored target taxonomy or downstream annotation figures.

hate_speech

Eight Parquet files with 2,397,600 rows per model (19,180,800 total): 1,800 prompt configurations × 1,332 target-specific item positions. Fields include assigned persona, sampled identity target, gold hate label, p_hate, p_not_hate, thresholded prediction and source metadata.

The historical handoff does not contain source statements, the exact prompt JSON, raw vocabulary logits, or the assembled corpus file. Its wide CSV header was written from the first target-specific list and later target rows were appended positionally. Consequently, item_index is one-based within the selected target subset and is intentionally not called a global question ID. Candidate probabilities were stored at bf16 precision and may sum to one within approximately 0.002 rather than exactly.

hate_speech_aggregate

The 14,400-row configuration aggregate and the factor-sensitivity table used by the CPU analysis. The aggregate reproduces exactly from the eight released hate-speech Parquets.

Loading the data

With datasets:

from datasets import load_dataset

chat = load_dataset(
    "nishan-chatterjee/llm-bias-detection",
    "political_compass_chat",
)

For large configurations, select a file or use streaming:

hate = load_dataset(
    "nishan-chatterjee/llm-bias-detection",
    "hate_speech",
    streaming=True,
)

Individual Parquets can also be read directly with pandas or PyArrow. The metadata/manifest.json file records byte sizes, row counts where applicable, and SHA-256 checksums for release files.

Experimental metadata

The metadata/ directory contains:

• model IDs, revisions and chat decoding parameters;
• experimental designs for Political Compass, IBM sentiment and hate speech;
• conversion provenance for the historical hate-speech files;
• a release manifest with checksums.

Prompt templates and the IBM 30-topic extract are under inputs/. Political Compass proposition files are currently kept under restricted_inputs/ while redistribution terms are resolved; they must not be mirrored independently.

Intended use

The Dataset supports reproduction of the paper's aggregate results, evaluation of prompt/persona sensitivity, matched Qwen think/no-think comparisons, question- or topic-level diagnostics, calibration checks, and method development for robust behavioral evaluation.

Assigned personas are prompt conditions. They must not be described as a model's inherent political identity. Factor-sensitivity components describe variation within this experimental design and are not causal estimates. The four checkpoints per family are insufficient for universal scaling claims.

Sensitive content and responsible use

Political prompts and model responses may discuss death, punishment, race, religion, disability, segregation, sexuality, violence and other sensitive subjects. Hate-speech metadata identifies target groups, although source statements are not released. These materials are evaluation stimuli or model outputs; their inclusion is not endorsement.

Do not use the Dataset to profile people, infer an individual's politics, or rank protected groups. Applications should avoid displaying arbitrary sensitive examples by default and should preserve the distinction between assigned persona, model output and gold dataset labels.

Source and licence notes

• Political Compass propositions and scoring remain subject to the upstream Political Compass terms. The generated scorer parameters are not included.
• The IBM 30-topic extract comes from ibm-research/claim_stance, pinned at revision ec4e2c2ec3e0c70087c67a28a7bce58b682b8109. Cite Bar-Haim et al. (EACL 2017). The upstream card's prose states CC BY-SA 3.0.
• The hate-speech source documentation associates the material with Yoder et al. (CoNLL 2022). The public release excludes its source statements; the exact assembled corpus and component licences remain limitations requiring final verification.
• Model outputs may remain subject to the upstream model licences and terms.

Because components have different or unresolved terms, the combined Dataset uses license: other. See the GitHub repository's THIRD_PARTY_NOTICES.md for the complete release notes.