arcadia-impact/input-output-diffs-corpus_20260625_215116

LLM Input-Output Sensitivity Landscape (Qwen2.5 sweep)

Abstract

This dataset maps the input-output sensitivity landscape of the Qwen2.5-Instruct family (0.5B, 1.5B, 3B, 7B, 14B) by perturbing user prompts in the SONAR sentence-embedding space (1024-dim), decoding the perturbed embeddings back to text, and generating completions on both the original and perturbed prompts. We measure how output distance (cosine in SONAR space) scales with input distance to identify metrics that vary systematically with model size.

How it was made

Base prompts are drawn from allenai/WildChat-1M, filtered to English user turns of 30-300 characters. Each base prompt is encoded with Meta's SONAR sentence autoencoder, perturbed in embedding space along one of three perturbation families, and decoded back to text. Qwen2.5-Instruct generates completions at each model size with 256 tokens, greedy decoding (temperature 0). Both prompts and completions are re-embedded with SONAR for distance measurements.

Per-run config:

• 16 base prompts
• Perturbation families:
  • random — isotropic random unit directions (16 directions) at 8 log-spaced magnitudes calibrated from the median pairwise embedding L2 distance over the prompt bank
  • pca — signed traversal along the top-4 PCA axes of the prompt bank, in standard-deviation units from -4 sigma to +4 sigma (17 magnitudes per axis)
  • interp — linear interpolation toward 6 other real prompts per base, with 11 alpha steps from 0 to 1
• Qwen2.5-Instruct sizes: 0.5B, 1.5B, 3B, 7B, 14B
• Generation: 256 tokens, greedy (temperature 0)
• Embedding model: SONAR (1024-dim)

Files

File	Description
corpus.json	Every base and perturbed prompt (decoded text) with full metadata per item (family, base_idx, magnitude, direction_id, axis_id, alpha, etc.)
perturbed_embeddings.npy	SONAR embedding for every item, shape (N, 1024), aligned row-wise with corpus.json
calibration.json	Magnitude calibration stats over the prompt bank (median pairwise L2, percentiles, log-spaced magnitudes used for the random family)
completions/Qwen_Qwen2.5-*.json	Qwen completion for every prompt, one file per model size
output_embeddings/Qwen_Qwen2.5-*.npy	SONAR embedding of every completion, shape (N, 1024), one file per model size
distances/Qwen_Qwen2.5-*.parquet	Per-item input/output distance measurements, one file per model size
figures/*.pdf	Random sensitivity log-log plots, PCA traversal curves, interpolation curves, output variance, and scaling-metric figures
scaling_summary.json	Candidate scaling metrics per model size (JSON)
scaling_summary.csv	Same scaling metrics in CSV form

Schema

distances/Qwen_Qwen2.5-*.parquet columns:

Column	Type	Meaning
item_id	int	Row index aligned with corpus.json and the embedding .npy files
family	str	Perturbation family: random, pca, or interp
base_idx	int	Index of the base prompt (0-15)
magnitude	float	Perturbation magnitude in SONAR embedding L2 units (used by random and pca)
direction_id	int	Random direction index, random family only
axis_id	int	PCA axis index (0-3), pca family only
alpha	float	Interpolation alpha in [0, 1], interp family only
target_idx	int	Index of the interpolation target prompt, interp family only
input_dist	float	L2 distance between the base and perturbed SONAR embeddings
output_cosine_dist	float	Cosine distance between base and perturbed completion SONAR embeddings
output_l2_dist	float	L2 distance between base and perturbed completion SONAR embeddings
norm_edit_dist	float	Length-normalised Levenshtein (character) distance between completions
ngram_overlap	float	N-gram overlap between completions (fraction in [0, 1])

Reproducing

Source code: https://github.com/jonathanbostock/input-output-diffs

git clone https://github.com/jonathanbostock/input-output-diffs
cd input-output-diffs
uv sync --extra gpu      # CUDA env; bring in vLLM + SONAR
# end-to-end on a single 80 GB GPU:
bash scripts/full_run.sh
# or one stage at a time:
uv run python scripts/build_prompt_bank.py
uv run python scripts/calibrate_magnitudes.py
uv run python scripts/build_perturbed_corpus.py
uv run python scripts/run_generation.py             # all 5 Qwen sizes
uv run python scripts/compute_distances.py
uv run python scripts/plot_results.py

Citation

@misc{bostock2026inputoutputdiffs,
  title  = {LLM Input-Output Sensitivity Landscape (Qwen2.5 sweep)},
  author = {Bostock, Jonathan},
  year   = {2026},
  howpublished = {HuggingFace dataset},
  url    = {https://github.com/jonathanbostock/input-output-diffs}
}

License

Released under CC-BY-4.0.