migrate-activations

Migrate stored activation data from the Supabase database to the wisent-ai/activations HuggingFace dataset. Organized as sub-actions for creating the repo, migrating in bulk or per combo, building the index, verifying and cleanup.

Basic Usage
python -m wisent migrate-activations ACTION [OPTIONS]

Examples

Create the dataset repo
python -m wisent migrate-activations create-repo
Migrate a single combo
python -m wisent migrate-activations single   --model meta-llama/Llama-3.1-8B-Instruct   --benchmark truthfulqa_mc1   --strategy chat_last
Migrate everything in a slice
python -m wisent migrate-activations all   --combo-start 0 --combo-end 100

Arguments

Actions

ArgumentDescription
create-repoCreate the wisent-ai/activations HF dataset repo
allMigrate all activations from Supabase to HF
consolidateBuild a unified index.json from marker files
singleMigrate a single model/benchmark/strategy combo
pair-textsMigrate pair texts for a benchmark
rawMigrate raw activations (chunked)
verify-completenessCheck all Supabase combos exist in the HF index
cleanupDrop Supabase rows for combos already in HF
verifyVerify HF data matches Supabase

Common Options

ArgumentDefaultDescription
--database-urlDATABASE_URL envDatabase URL
--model / --benchmark / --strategyrequiredCombo selectors for single, raw and verify

Upload sub-actions accept a shared set of HuggingFace retry flags (--hf-max-retries, --hf-base-wait, --hf-backoff-max-exp, --hf-jitter-min, --hf-jitter-max, --hf-retryable-pattern) and a preview flag that reports what would be uploaded without writing; run wisent migrate-activations <action> --help for the per-action arguments.

Related Commands

  • get-activations - Collect activations from contrastive pairs
  • zwiad - Geometry analysis over stored activations
  • tasks - Run the evaluation pipeline that fills the database

Stay in the loop. Never miss out.

Subscribe to our newsletter and unlock Wisent insights.