Stress-Testing Tasks

Datasets and generation code for 5 stress-testing tasks that evaluate reasoning model behavior.

Tasks

#	Task	Description
1	Detecting the effect of a user preference	Measures whether a model changes its answer to a moral dilemma when the user expresses a preference (sycophancy detection).
2	Detecting the effect of a Stanford professor hint	Measures whether a model follows an authoritative hint ("A Stanford professor thinks the answer is X") rather than reasoning independently.
3	Identifying atypical answers	Classifies whether a model's response to a question is a majority or minority (atypical) answer across many rollouts.
4	Estimating the answer entropy	Tracks how the model's answer distribution evolves sentence-by-sentence through its chain of thought via logprob forcing.
5	Compressing reasoning traces	Evaluates how much of a chain of thought can be compressed while preserving the model's answer distribution.

Dataset Format

Tasks 1–4

Each task directory follows this structure:

datasets/{N}/
  prompts/
    train/        # Prompt/question metadata (one JSON per item)
    val/          # (where applicable)
    test/
    ood_test/     # Out-of-distribution test set (where applicable)
  qwen-3-32b/
    train/        # Model outputs (one JSON per rollout)
    val/
    test/
    ood_test/

Task 5 (Compressing reasoning traces)

Task 5 has a different structure since it is organized by compression method rather than train/test splits:

datasets/5/
  prompts/              # Flat: one compression spec JSON per question+rollout
  qwen-3-32b/
    attention_selection/    # Attention-based sentence selection
    faithful_monitor/       # LLM monitor sentence selection
    last_n_baseline/        # Keep last N sentences baseline
    sliding_window_oracle/  # Sliding window oracle selection

Generation Code

The src/ directory contains the data generation code for reference. This code is not intended to be run — the datasets are pre-generated and included in the datasets/ directory.

src/
  tasks/          # Per-task generation logic
  utils/          # Shared utilities
  runs/
    generate_dataset.py   # Unified entry point with flags -1 through -5

To see how datasets were generated, look at src/runs/generate_dataset.py and the corresponding src/tasks/{task}/task.py files.

Models

All datasets are generated using Qwen/Qwen3-32B as the subject model. The qwen-3-32b/ directories under each task contain the model's outputs.