Adam Qwen3-8B AO Reupload: LatentQA + Classification + Past Lens

What This Checkpoint Is

• Source repo: adamkarvonen/checkpoints_latentqa_cls_past_lens_addition_Qwen3-8B
• Base model: Qwen/Qwen3-8B
• Adapter format: PEFT LoRA
• Hook layer: 1
• Activation readout layer options in ao_config.json: [9], [18], [27] corresponding to 25/50/75% depth
• Seed: 42
• Learning rate: 1e-5
• Train batch size: 16
• Epochs: 1
• Paper shorthand token label: ~60M

Best-Available Training Mixture

• past_lens single-activation loader: num_train=100000, max_k_activations=1, directions=[past,future], max_length=512
• past_lens multi-activation loader: num_train=100000, max_k_activations=50, directions=[past,future], max_length=512
• latentqa: num_train=100000, max_window_size=3, position_types=[all,window]
• classification families, each with a single-token loader (max_window_size=1) and a multi-token loader (max_window_size=50):
  • geometry_of_truth
  • relations
  • sst2
  • md_gender
  • snli
  • ag_news
  • ner
  • tense
  • language_identification
  • singular_plural (listed with num_train=0 in the checkpoint config)

Documented Aggregate Stats

• total samples: 1,027,328
• total tokens: 66,469,521
• past_lens: 584,488 samples, 42,003,254 tokens
• classification (all): 378,000 samples, 18,407,443 tokens
• latentqa (all): 64,840 samples, 6,058,824 tokens

Notes

• The checkpoint card is based on the source repo's ao_config.json plus Adam's local training-data counting script in the Activation Oracles reference code.
• This repo is a mirror of the adapter weights and config, with only the README replaced.