solarkyle/jspace-lenses

jspace lenses: fitted Jacobian lenses for 5 open models

Code (this card's jlens snippets, the sidecar, all analysis) is Apache 2.0. The Gemma lens files are derived from Gemma model weights and are provided subject to the Gemma Terms of Use; the Qwen lens is derived from an Apache 2.0 model. Using a Gemma lens requires the Gemma model itself, which you accept the license for on its own HF page.

Fitted Jacobian lenses ("J-lens", from Anthropic's July 2026 global-workspace paper and the anthropics/jacobian-lens reference implementation, Apache 2.0) for five open-weight models, plus the raw evaluation traces behind the findings at github.com/solarkyle/jspace.

Path	Model	Fit prompts	HF revision at fit time
gemma-4-e4b-it/lens.pt	google/gemma-4-E4B-it	100	fee6332c1aba
gemma-4-12b-it/lens.pt	google/gemma-4-12B-it	75	5926caa4ec0c
huihui-gemma-4-12b-it-abliterated/lens.pt	huihui-ai/Huihui-gemma-4-12B-it-abliterated	75	060ea173c4d1
gemma-4-26b-a4b-it/lens.pt	google/gemma-4-26B-A4B-it (MoE)	100	20da991ab4af
qwen3.6-27b/lens.pt	Qwen/Qwen3.6-27B	100	6a9e13bd6fc8

All fits: WikiText-103 prompts, bf16, dim-sharded and merged with JacobianLens.merge (details and the consumer-GPU recipe in the repo).

Load one

import jlens, transformers, torch
from huggingface_hub import hf_hub_download

path = hf_hub_download("solarkyle/jspace-lenses", "gemma-4-e4b-it/lens.pt")
lens = jlens.JacobianLens.load(path)

tok = transformers.AutoTokenizer.from_pretrained("google/gemma-4-E4B-it")
hf = transformers.AutoModelForImageTextToText.from_pretrained(
    "google/gemma-4-E4B-it", dtype=torch.bfloat16, device_map="cuda")
model = jlens.from_hf(hf, tok)
logits, _, _ = lens.apply(model, "Fact: the currency of the country shaped like a boot is", positions=[-1])

Eval traces (traces/)

• uncertainty_trivia_<model>.jsonl - 500 TriviaQA questions per model: answer, correctness, output-confidence baselines, workspace features, per-layer entropy trajectory. Basis of the hallucination-prediction results.
• uncertainty_fake_<model>.jsonl - 100 matched real/fabricated-entity questions per model.
• uncertainty_shape_<model>.jsonl - the noise-decomposition traces (rival mass vs tail smear) behind the "entropy = tail smear" result; uncertainty_shapeq4_* are the same protocol at NF4 4-bit (the lens-survives-quantization check).
• emotion_matrix_5models.json - covert-emotion delta matrices + evidence.
• workspace_dump.json - top workspace tokens per layer under 6 covert emotional states, all 5 models.

Findings, caveats, and every analysis script: github.com/solarkyle/jspace

Router weights (router/)

workspace_router_e4b.json - the trained hallucination-risk classifier (logistic regression, 10 workspace features / 14 combined) from the repo's analyze_router.py. Trained on E4B, transfers zero-shot to the other Gemmas when features are z-scored per model (AUC 0.74-0.78). Predicts P(wrong); escalate when risk exceeds your budget threshold. Feature definitions in analyze_router.py.