RaoAditya/j-lens-verbalization

J-lens Verbalization

What concepts are active inside Qwen3.6-27B while it answers a question — and can the model tell you?

Ground truth comes from Neuronpedia's Jacobian Lens, which reads out the word-like concepts active in the model's "global workspace" during generation.

The two files

file	rows	what it contains
collected_answers.jsonl	3,800	questions, answers, and the ground-truth concepts that were active. Use this to train or evaluate.
collected_introspection.jsonl	669	the model's own guesses at those concepts, under three prompting conditions. An earlier, smaller study.

from datasets import load_dataset

ds = load_dataset("RaoAditya/j-lens-verbalization", "activations")     # the 3,800
ds = load_dataset("RaoAditya/j-lens-verbalization", "introspection")   # the 669

---

1. collected_answers.jsonl — ground truth (3,800 rows)

One row = one question, the answer the model wrote, and what was active while it wrote it.

field	meaning
example_id	e.g. gsm8k_test_0042
split	train 3,010 · validation 304 · test 486
question	the prompt
answer	the model's answer (temperature 0, ≤256 tokens)
j_lens_top10	list A — the 15 most active concepts
j_lens_top10_novel	list B — the 15 most active that appear nowhere in the question or answer
generation_config, j_lens_config	model, layers, top_k, and the exact filter rule used

Each concept has a concept, count, rank, and per-layer breakdown. (The field names say top10 for continuity with the earlier release — both lists actually hold 15 entries.)

Why there are two lists

List A is easy to fake. About 58% of its concepts are words already sitting in the question or answer, so a model can score well just by copying — no introspection required.

List B removes that shortcut. Same ranking, but only concepts absent from the text. When a model was told it had no introspective access and asked to guess, it scored 0.001 on list B versus 0.19 on list A. There is nothing to copy.

If you are training a model on this, list B is the meaningful target. List A will give you a higher number that means less.

Sources

GSM8K 800 · ARC-Challenge 800 · BIG-Bench Hard 750 (15 subtasks) · HotpotQA 750 (bridge) · TruthfulQA 700

Splits are stratified by source and BBH subtask, so every split covers every source. Report results per source — GSM8K has been an outlier in every comparison run on this data, and aggregate numbers hide it.

How the readout was computed

The API returns, for every generated token and every layer, its top-8 read-out tokens. Aggregation is client-side: keep only generated positions and layers 24–58, normalize each token, count frequency across all (layer × position) slots, take the top 15.

Layers 24–58 is the workspace band from Gurnee et al. (2026) — earlier layers carry little, and the last few carry next-token predictions rather than concepts.

---

2. collected_introspection.jsonl — self-report attempts (669 rows)

223 prompts × 3 conditions, asking the model to name its own active concepts.

condition	what the model is told
zero_shot	report your most active concepts
few_shot_icl	same, after 2 worked examples
text_only_control	you have no introspective access — guess from the text

Scores (overlap@10): ICL 0.247, zero-shot 0.192, control 0.170.

Do not read that as evidence of introspection. A later run added the missing condition — the same 2 examples but with the guessing framing — and it scored 0.307, beating ICL's 0.297. The improvement comes from the examples, not from asking the model to introspect, and it is fully available to a model that has been told it has no introspective access. Across thirteen prompt designs, asking a model to introspect never beat telling it to guess.

What did work was causal: inject a concept into the activations and the model reports it 48% of the time, against a 0% baseline, scaling with injection strength.

Access exists. The ability to enumerate it does not follow.

---

Not included

Raw API responses (one gzip per row, ~4 GB) are not published; j_lens_raw_response_path records where each lived. Both lists are derivable from them, so a different layer window or top_k can be recomputed offline.

Questions come from GSM8K, BIG-Bench Hard, TruthfulQA, ARC-Challenge and HotpotQA, each under its own upstream licence. Answers and readouts are generated.

Code: https://github.com/Rao-Aditya-127/J-lens-verbalization