Downloads · 30 days
79
20% of all-time downloads
ceselder/adam-reupload-qwen3-8b-latentqa-cls-past-lens
adam-reupload-qwen3-8b-latentqa-cls-past-lens is a text generation model from ceselder. Use it when you need the model to write or continue text. It is set up for peft.
This repo is a ceselder re-upload of Adam Karvonen's checkpointslatentqaclspastlensadditionQwen3-8B activation-oracle adapter so the collection can carry a detailed model card.
Downloads · 30 days
79
20% of all-time downloads
All-time downloads
405
Public
Repo size
698 MB
Likes
0
Public
Click a slice to open those files.
.safetensors698 MB · 100%
From the Hugging Face model README
This repo is a ceselder re-upload of Adam Karvonen's checkpoints_latentqa_cls_past_lens_addition_Qwen3-8B activation-oracle adapter so the collection can carry a detailed model card.
adamkarvonen/checkpoints_latentqa_cls_past_lens_addition_Qwen3-8BQwen/Qwen3-8B1ao_config.json: [9], [18], [27] corresponding to 25/50/75% depth421e-5161~60MThis checkpoint's own ao_config.json lists the following dataset loaders:
past_lens single-activation loader: num_train=100000, max_k_activations=1, directions=[past,future], max_length=512past_lens multi-activation loader: num_train=100000, max_k_activations=50, directions=[past,future], max_length=512latentqa: num_train=100000, max_window_size=3, position_types=[all,window]max_window_size=1) and a multi-token loader (max_window_size=50):
geometry_of_truthrelationssst2md_gendersnliag_newsnertenselanguage_identificationsingular_plural (listed with num_train=0 in the checkpoint config)Adam's local count_training_data.py comments document the closest default mixture as:
1,027,32866,469,521past_lens: 584,488 samples, 42,003,254 tokensclassification (all): 378,000 samples, 18,407,443 tokenslatentqa (all): 64,840 samples, 6,058,824 tokensao_config.json plus Adam's local training-data counting script in the Activation Oracles reference code.