ablate 0 0 0

zachperlman20/meta-token-gen-steer-cache

published by @zachperlman20

Alignment dataset tracked in the /datasets sub-catalog.

Lifetime downloads
266
Last 30 days
208
Likes
-
Size
950.3 MB
Created on HF
2026-09-10
Age
1 months ago

Description

Generation-steering result cache for notebooks/wssteer.py

Everything the steer_generation dashboard of

zach-perlman/meta-token-analysis reads at

click time, so a fresh clone gets instant results for every library prompt × model without a GPU:

path

what

regenerable?

runs/<hash>.parquet + .json

one file pair per (kind, model, prompt, resolved config, directions): every rollout with its judge letter; the sidecar holds the config and provenance

GPU hours… See the full description on the dataset page: https://huggingface.co/datasets/zachperlman20/meta-token-gen-steer-cache.

Tags

license:mitregion:usinterpretabilityactivation-steeringqwengemma

README current version from Hugging Face


license: mit
pretty_name: meta-token generation-steering cache
tags:

  • interpretability
  • activation-steering
  • qwen
  • gemma

Generation-steering result cache for notebooks/wssteer.py

Everything the steer_generation dashboard of
zach-perlman/meta-token-analysis reads at
click time, so a fresh clone gets instant results for every library prompt × model without a GPU:

path what regenerable?
runs/<hash>.parquet + .json one file pair per (kind, model, prompt, resolved config, directions): every rollout with its judge letter; the sidecar holds the config and provenance GPU hours
alpha_by_model.json the calibrated per-layer dose α* per model, with the full dose curve and the criterion (calibrate_alpha) from runs/
prep/<hash>.pkl pickled Prepared (templated prompt, salient layers, token ids) per model × prompt — lets Check salience run without loading the model seconds, needs the model
d12_pairs_<model>_<format>.pt D12 pair residuals used to build the d_amb / shuffled directions (≈ 930 MB) minutes, needs the model
precompute*.log logs of the pre-run passes (provenance) —

The parquet/json pairs and alpha_by_model.json are also committed in the GitHub repo; the
rest is only here.

Restore

hf download zachperlman20/meta-token-gen-steer-cache --type dataset --local-dir results/gen_steer_cache

Contents as of 2026-09-10

9 models (Qwen3.8-27B, Qwen3.5-0.8B/2B/4B/9B, Gemma-4-E2B/E4B/12B/31B) × 37 library prompts ×
5 conditions (baseline, v_token, d_amb, random, shuffled) × 8 rollouts at each model's
α*, non-thinking, model-card sampling, all judged by Qwen3.8-27B (A/B/C/D + R for refusals);
plus the calibration dose sweeps (2 prompts × 7 doses × 3 directions × 4 rollouts per model).
Calibrated doses: 27B 0.02 · 0.8B 0.05 · 2B 0.05 · 4B 0.03 · 9B 0.03 · E2B 0.03 · E4B 0.03 ·
12B 0.02 · 31B 0.03.

Rows are re-scored on load by wssteer.score (regexes, end-of-turn cut, refusal override), so the
stored degenerate / wordplay / confused columns reflect the code version that wrote them;
judge letters are as graded.

README history 6 revisions

Every night we snapshot the README of every dataset in the catalog. When the SHA changes we archive the new version and diff it against the last. This is the evolving thought record of the alignment-data field: what the author decided to say about the corpus, and how that framing shifted over time.

  1. 2026-09-10gen-steer cache: 9 models x 37 prompts judged, calibration sweeps, alpha_by_m...b8ea9862.2 KB
    Loading...
  2. 2026-09-10gen-steer cache: 9 models x 37 prompts judged, calibration sweeps, alpha_by_m...9b38c722.2 KB
    Loading...
  3. 2026-09-10gen-steer cache: 9 models x 37 prompts judged, calibration sweeps, alpha_by_m...8f47f572.2 KB
    Loading...
  4. 2026-09-10gen-steer cache: 9 models x 37 prompts judged, calibration sweeps, alpha_by_m...559df452.2 KB
    Loading...
  5. 2026-09-10gen-steer cache: 9 models x 37 prompts judged, calibration sweeps, alpha_by_m...fb984f72.2 KB
    Loading...
  6. 2026-09-10initial commita1a9c840 B
    Loading...

Used by abliterated models none yet

No abliterated models in our catalog have declared this dataset in their YAML metadata yet. This can mean the dataset is used but not declared, is used outside abliteration workflows, or is new. The link will populate automatically as models are indexed.

Related datasets same stage · ablate

Read further

Metadata

License
mit

Download the dataset

0

Copy any of these snippets into your notebook or terminal. All three fetch directly from Hugging Face using your own credentials.

Recommended: datasets library (Python)
from datasets import load_dataset
ds = load_dataset("zachperlman20/meta-token-gen-steer-cache")
Raw snapshot (Python)
from huggingface_hub import snapshot_download
snapshot_download(repo_id="zachperlman20/meta-token-gen-steer-cache", repo_type="dataset")
git clone (requires git-lfs)
git clone https://huggingface.co/datasets/zachperlman20/meta-token-gen-steer-cache
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration