Skip to content

Add LatentQA and Activation Oracle reading methods#150

Open
aryamanarora wants to merge 1 commit into
stanfordnlp:mainfrom
aryamanarora:latentqa-reading
Open

Add LatentQA and Activation Oracle reading methods#150
aryamanarora wants to merge 1 commit into
stanfordnlp:mainfrom
aryamanarora:latentqa-reading

Conversation

@aryamanarora

Copy link
Copy Markdown
Member

Summary

  • Adds LatentQAReading and ActivationOracleReading model classes for concept detection via external decoder models (Yes/No logit scoring + rating variants)
  • Includes data prep scripts (prepare_data.py, setup-latentqa.sh) and sweep configs for Llama-3-8B, Llama-3.1-8B, and Qwen3-8B
  • Infra changes to inference.py, evaluate.py, train.py to support these in the latent evaluation pipeline
  • Steering variants are not included — this is reading/detection only

Setup

# Clone external repos into axbench/models/
cd axbench/models && git clone https://github.com/aypan17/latentqa.git _latentqa
cd axbench/models && git clone https://github.com/adamkarvonen/activation_oracles.git _activation_oracles

Test plan

  • Verify from axbench.models.latentqa import LatentQAReading works after cloning external repos
  • Verify from axbench.models.activation_oracle import ActivationOracleReading works
  • Run latent inference with reading_llama3_8b.yaml config
  • Run latent evaluation on inference results

🤖 Generated with Claude Code

Adds two new model classes for concept detection via external decoder models:
- LatentQAReading / LatentQAReadingRating: detect concepts using the LatentQA decoder
- ActivationOracleReading / ActivationOracleReadingRating: detect concepts using activation oracle LoRA adapters

Includes data prep scripts, sweep configs, and infra changes to support
these in the inference/evaluation pipeline. Steering variants are excluded.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant