sida/icalens-qwen3.5-2b-ultrachat-1m

ICA Lens for Qwen/Qwen3.5-2B

This repository contains a fitted ICA Lens for analyzing internal activations of Qwen/Qwen3.5-2B. It provides layer-wise ICA transformations for mapping residual-stream activations to independent-component scores and energy shares.

Artifact summary

Field	Value
Analyzed model	Qwen/Qwen3.5-2B
Analyzed model revision	15852e8c16360a2fea060d615a32b45270f8a8fc
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
Hidden size	2048
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	10000000
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:

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 reuse the R-lens fitted for base model Qwen/Qwen3.5-2B-Base. 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	2048	1000000	50
1	2048	1000000	50
2	2048	1000000	50
3	2048	1000000	50
4	2048	1000000	50
5	2048	1000000	50
6	2048	1000000	50
7	2048	1000000	50
8	2048	1000000	50
9	2048	1000000	50
10	2048	1000000	50
11	2048	1000000	50
12	2048	1000000	50
13	2048	1000000	50
14	2048	1000000	50
15	2048	1000000	50
16	2048	1000000	50
17	2048	1000000	50
18	2048	1000000	50
19	2048	1000000	50
20	2048	1000000	50
21	2048	1000000	50
22	2048	1000000	50
23	2048	1000000	50

Fitting provenance

{
  "activation_dataset": {
    "dtype": "bfloat16",
    "format": "icalens.activations",
    "format_version": 1,
    "manifest_sha256": "3336b0c44c11509299a2436a20695f848026599102f46e586afa3ebd28acc61f"
  },
  "candidate_tokens": 10000000,
  "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

@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}
}