wikitext dict | multihop dict | published_36_on_36 dict |
|---|---|---|
{
"4": [
50885,
454
],
"8": [
60810,
353
],
"12": [
19582,
356
],
"16": [
5486,
477
],
"20": [
962,
744
],
"24": [
852,
654
],
"28": [
528,
624
],
"32": [
468,
601
],
"36": [
515,
1133
],
"40": [
600,
6... | {
"4": [
26197,
22642
],
"8": [
84539,
39620
],
"12": [
7660,
51138
],
"16": [
5042,
38980
],
"20": [
977,
26643
],
"24": [
121,
19219
],
"28": [
88,
25003
],
"32": [
570,
29185
],
"36": [
24,
8258
],
"40": [
... | {
"wikitext_jlens": {
"8": 182,
"16": 106,
"24": 61,
"32": 47
},
"wikitext_logitlens": {
"8": 11793,
"16": 42247,
"24": 28634,
"32": 672
}
} |
{
"4": [
41377,
399
],
"8": [
43645,
349
],
"12": [
12916,
367
],
"16": [
3798,
413
],
"20": [
2707,
721
],
"24": [
1339,
832
],
"28": [
992,
814
],
"32": [
719,
791
],
"36": [
654,
1747
],
"40": [
881,
... | {
"4": [
20586,
28912
],
"8": [
63868,
54560
],
"12": [
17846,
70204
],
"16": [
5739,
70285
],
"20": [
1015,
45682
],
"24": [
38,
28224
],
"28": [
51,
38115
],
"32": [
470,
44822
],
"36": [
17,
6273
],
"40": [
... | {
"wikitext_jlens": {
"8": 182,
"16": 106,
"24": 61,
"32": 47
},
"wikitext_logitlens": {
"8": 11793,
"16": 42247,
"24": 28634,
"32": 672
}
} |
Jacobian Lens Transfer: Qwen3.6-27B → Qwen3.8-27B
Evaluation data and code for the preprint "Survival of the Fitted: Cross-Checkpoint Transfer of a Jacobian Lens from Qwen3.6-27B to Qwen3.8-27B" (Shorthill, 2026, v1.1; Zenodo DOI 10.5281/zenodo.22133736).
A published Jacobian lens fitted on Qwen3.6-27B
(neuronpedia/jacobian-lens,
revision qwen-n1000) is applied unmodified to Qwen3.8-27B. Tuned-lens
translators are known to transfer from a base model to its fine-tuned
variants; the open question here is whether a Jacobian lens preserves its
latent semantic readouts and its causal directions across a model-version
change. Under one protocol on both models: the latent-entity readout (two-hop
unstated middle entity, 40 prompts) survives the transfer with a
depth-dependent cost. Median target rank rises from 4 to 17 at layer 48
(1.7–3.4× paired rank penalty at layers 44–48, p < 10⁻³), sits at parity at
layer 40, and is better on the successor at layers 24–36 (38 vs 121 at
layer 24; paired sign tests, p < 10⁻³). WikiText next-token readout shows a
distinct, depth-dependent transfer cost: 1.2–1.3× paired rank inflation at
mid-network layers and about 2× by layer 48; the raw logit-lens ranks
shift by a similar 1.2–1.3× at mid-network layers, so the lens-specific
cost is concentrated near layer 48. Causal directions transfer in the one case
tested: ablating a paradox-direction bundle (pullbacks of " paradox",
" paradoxical", 悖论, 矛盾; layers 18–47) removes the paradox readout and the
word from generation in all 8 cells on both models, with reasoning mode
enabled and disabled. Readout identity is stable under temperature-0.7
sampling. Transfer is therefore informative but not calibration-preserving;
see the preprint for the full statistics and limitations.
Files
Rank evaluations (reading)
eval38_big.json,eval36_big.json— per-layer median ranks (J-lens and raw logit lens): 100 WikiText-103 passages ×7 positions, 40 two-hop prompts. These are the runs behind every number in the paper.raw38.json,raw36.json— full per-prompt/per-sample rank lists behind the paired statistics (sign tests, bootstrap CIs).eval38_big.py,eval36_big.py,raw38.py,raw36.py— the exact evaluation scripts (require thejlensreference implementation).mh_items.py— the 40-prompt two-hop set.
Causal check (steering)
escher_ablation/— 8 cells per model:{36,38}_{think,direct}_{clean,ablate}.json+_field.npz(full top-256 transported field at all lens layers). Prompt: "Describe Escher's impossible staircase"; ablation projects out the Gram–Schmidt orthogonalized paradox bundle at layers 18–47 during generation.escher_ablate.py— the ablation script.
Temperature check (sampling stability)
temperature_check/— pairs p01 (truth-held) and p02 (verbatim reuse), arms A/B:*_greedy.json/*_greedy_field.npzare the temperature-0 baselines;*_t07.json/*_t07_field.npzare the temperature-0.7 runs.ab_temp07.py— the sampling-run script.
Figure capture
figure_capture/— prefill-only full-field capture (top-256, all lens layers) of the flagship two-hop prompt on both models, used for the readout-board figure in the paper, with the capture script.
Exploratory
hum38.py,hum38_run.json,hum38_field.npz— fixed introspective prompt+prefill on Qwen3.8-27B with the full top-256 transported field at all 63 lens layers (indicesint32,probsfloat16,layers; positions cover prompt+generation).
Exploratory 35B follow-up (paper §2.4, §3.3, Table 2)
exploratory_35b/— the Qwen3.6-35B-A3B lens (stanleytheli, n=1000) applied unchanged to Qwen3.5-35B-A3B Instruct (BF16 GGUF) and to the HauhauCS Qwen3.6-35B-A3B abliterated derivative (Q8_K_P): per-target S1/S2 result files, the 40-prompt set, the evaluation driver and decoding code, the patched llama.cpp capture example and build script, run logs, andPROVENANCE.md(instrument, targets, runtime, recorded hash prefixes). Latent-entity evaluation only; provisional, see the paper.
Provenance
Models: Qwen/Qwen3.6-27B, Qwen/Qwen3.8-27B, bf16, greedy unless noted, single seed. Hardware: one RTX PRO 6000 Blackwell (96 GB). All runs 2026-08-14. See the preprint for protocol details, statistics, and limitations.
- Downloads last month
- 368