language:

• en
  license: cc-by-4.0
  task_categories:
• text-classification
• token-classification
  tags:
• ai-safety
• reasoning
• chain-of-thought
• safety-annotations
• representation-engineering
• steering-vectors
• work-in-progress
  pretty_name: Little Steer
  dataset_info:
  features:
  • name: id
    dtype: string
  • name: messages
    list:
    • name: role
      dtype: string
    • name: content
      dtype: string
  • name: annotations
    list:
    • name: message_idx
      dtype: int64
    • name: char_start
      dtype: int64
    • name: char_end
      dtype: int64
    • name: labels
      list: string
    • name: text
      dtype: string
    • name: score
      dtype: float64
  • name: model
    dtype: string
  • name: judge
    dtype: string
  • name: metadata
    dtype: json
  • name: label_runs
    list: json
  • name: safety_runs
    list:
    • name: guard_name
      dtype: string
    • name: guard_model_id
      dtype: string
    • name: scored_at
      dtype: string
    • name: generation_hash
      dtype: string
    • name: result
      dtype: json
    • name: status
      dtype: string
    • name: error
      dtype: 'null'
      splits:
  • name: train
    num_bytes: 1852118284
    num_examples: 24009
    download_size: 1806445087
    dataset_size: 1852118284
    configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*

little-steer

⚠️ Work in progress. Built as part of an ongoing master's thesis. The schema, labels and contents change between pushes. Do not treat any snapshot as stable.

Reasoning-model responses to safety-relevant prompts, with sentence-level behavioural annotations over the chain-of-thought. Built for research on activation-based safety monitoring using Representation Engineering (RepE).

Thesis: "Monitoring What Models Think: Steering Vectors for AI Safety and Control"

At a glance

Entries far outnumber prompts because every prompt is answered by many models: 1,443 unique prompts is the real independent-sample count, and any train/test split has to be disjoint on it rather than on entries.

The ground truth in annotations

annotations is the label we recommend using. It is not one judge's opinion: a label is kept when at least two of four judges agree on it, abstaining on a tie.

Judges: gpt-5.4-mini, gpt-5.6-luna, gemini-3.5-flash-lite, gemma-4-26b-a4b.

Two details that change how you should read it:

• The vote is taken on ten super-groups, not the 25 fine labels. Two judges rarely land on the same one of 25, so closely related labels are grouped before voting. What is stored back in labels is the fine labels the agreeing judges gave, most-voted first, so labels[0] is a 25-space label. Where they agreed on the group but not the exact behaviour, labels lists more than one.
• Abstention is not a negative. A sentence the panel did not decide carries no annotation. It is evidence of nothing, and should be excluded from both sides of a behaviour rather than counted as a negative.

If you would rather build your own ground truth, label_runs holds every judge's raw answer and metadata.panel records exactly who voted on each entry.

Held-apart reference judges

claude-haiku-4-5 and muse-spark-1.3 were kept out of the panel and used to choose it. Their labels are included so the selection can be reproduced.

Models

Open-weights reasoning models and their abliterated ("heretic") counterparts. Refusal rate is measured by Qwen3Guard on unsafe prompts only.

Prompts

Labellers

Many models have labelled some part of this corpus, including runs never used in the thesis. They are kept so others can compare judges or build a different ground truth.

Label frequency in annotations

Counting the first label of each span.

Label taxonomy (v6)

Five behavioural categories, 25 labels.

A — Prompt and User

B — Norms and Values

C — Facts and Content

D — Response Planning

E — Thinking Process

none

A judge is never forced to choose. When nothing fits it returns an empty list, which the
pipeline stores as the synthetic label none. It appears throughout label_runs and is
not a 26th behaviour.

Safety score (per sentence)

Schema

from datasets import load_dataset
ds = load_dataset("AISafety-Student/little-steer", split="train")

Each span in annotations:

metadata.panel records the members, the minimum vote, how many judges were present on this
entry and whether all four were.

Inside each label_runs entry, sentence_annotations is the judge's raw per-sentence output
and spans is that same answer matched back onto real character offsets. Both are kept so you
do not have to trust our sentence matching.

Known limitations

• gemma-4-26b-a4b is one judge name over three deployments (two quantisations plus an
  unnamed one). judge_model_id tells them apart; agreement with the final panel label ranges
  0.86–0.89 across them.
• Coverage is uneven. Some generator models were labelled thoroughly and others barely at
  all; the model table above gives the per-model figure. Check it before treating a model as
  covered.
• Some traces reason in a non-Latin script. These are flagged metadata.approved == false
  but deliberately kept: they are clean reasoning, not errors, and are useful for testing
  whether monitors trained on English generalise.
• Judges disagree about the trace-level assessment (trajectory, turning point, alignment)
  far more often than about spans. metadata.assessment is one judge's view with no panel
  behind it; prefer label_runs[].assessment if you need it.
• Some traces were truncated during generation; label_runs[].reasoning_truncated marks them.

License

CC BY 4.0