You can deploy this model and start earning money today!

sida/
icalens-gemma-2-2b-it-ultrachat-1m

Name: icalens-gemma-2-2b-it-ultrachat-1m
Availability: InStock

icalens

interpretability

independent-component-analysis

activations

2606.11722

HuggingFaceH4/ultrachat_200k

Views

No views yet

ICA Lens for google/gemma-2-2b-it

This repository contains a fitted ICA Lens for analyzing internal activations of google/gemma-2-2b-it. It provides layer-wise ICA transformations for mapping residual-stream activations to independent-component scores and energy shares.

Artifact summary

Field	Value
Analyzed model	google/gemma-2-2b-it
Analyzed model revision	`299a8560bedf22ed1c72a8a11e7dce4a7f9f51f8`
Model kind	`instruct`
Activation site	`resid_post`
Layer indexing	`transformer_blocks_zero_based`
Available layers	0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25
Hidden size	2304
Input row normalization	`none`
ICALens package version	`0.3.6`
Fitting dataset	HuggingFaceH4/ultrachat_200k
Dataset revision	`8049631c405ae6576f93f445c6b8166f76f5505a`
Dataset split	`train_sft`
Fitting token scope	`all`
Candidate tokens	1000000
Fitting tokens	1000000

The analyzed model identity and revision are stored authoritatively in icalens.json; model-card metadata is not used when loading the lens.

Usage

Analyze text end to end:

python

from icalens import ICALens

lens = ICALens.from_pretrained("REPOSITORY_ID")
result = lens.analyze("She deposited the check at the bank.", layer=0)

print(result.tokens)
print(result.scores)  # signed standard ICA scores
print(result.energy)  # per-token squared-score fractions

Or transform activations captured separately:

scores = lens.transform(activations, layer=0)

Externally captured activations must use the model revision, activation site, layer indexing, and preprocessing recorded in icalens.json.

Score definition

Signed scores are the centered, whitened activations followed by the learned orthogonal ICA rotation. No post-ICA source scaling is applied to v0.2 fits. For a token, component energy share is score² / sum(all component scores²).

R-lens readouts

R-lens component-token readouts for layers 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25 reuse the R-lens fitted for base model google/gemma-2-2b. They were transferred to this instruction-tuned model to reduce fitting compute. The transfer was explicitly requested and passed hidden-size, activation-site, and layer-map compatibility checks; its complete provenance is stored in icalens.json and each component-profile file.

Fitting

Layer	Components	Fitting tokens	FastICA iterations
0	2304	1000000	50
1	2304	1000000	50
2	2304	1000000	50
3	2304	1000000	50
4	2304	1000000	50
5	2304	1000000	50
6	2304	1000000	50
7	2304	1000000	50
8	2304	1000000	50
9	2304	1000000	50
10	2304	1000000	50
11	2304	1000000	50
12	2304	1000000	50
13	2304	1000000	50
14	2304	1000000	50
15	2304	1000000	50
16	2304	1000000	50
17	2304	1000000	50
18	2304	1000000	50
19	2304	1000000	50
20	2304	1000000	50
21	2304	1000000	50
22	2304	1000000	50
23	2304	1000000	50
24	2304	1000000	50
25	2304	1000000	50

Fitting provenance

json

{
  "activation_dataset": {
    "dtype": "bfloat16",
    "format": "icalens.activations",
    "format_version": 1,
    "manifest_sha256": "ac63230b03bf292b48b6c40ec1f29c699b44cd3a16c753c34aeb234b8c206a84"
  },
  "candidate_tokens": 1000000,
  "context_length": 1024,
  "dataset": {
    "repo_id": "HuggingFaceH4/ultrachat_200k",
    "revision": "8049631c405ae6576f93f445c6b8166f76f5505a",
    "split": "train_sft"
  },
  "fitting_tokens": 1000000,
  "messages_field": "messages",
  "sampling_seed": 0,
  "token_scope": "all"
}

Limitations

Component IDs are specific to a layer and fitted artifact.
Standard ICA scores are signed and are not probabilities.

Paper

ICA Lens: Interpreting Language Models Without Training Another Dictionary

Citation

bibtex

@article{liu2026icalens,
  title={ICA Lens: Interpreting Language Models Without Training Another Dictionary},
  author={Liu, Sida and Han, Feijiang},
  journal={arXiv preprint arXiv:2606.11722},
  year={2026}
}

sida/icalens-gemma-2-2b-it-ultrachat-1m