Downloads · 30 days
0
575-lab/kiji-inspector-Qwen-Qwen3.6-35B-A3B
kiji-inspector-Qwen-Qwen3.6-35B-A3B is a machine learning model from 575-lab. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for kiji-inspector. The card lists the license as apache-2.0.
This repository contains the sparse autoencoders (SAEs), feature labels, and direction-aware feature-ablation results produced by Kiji Inspector for Qwen/Qwen3.6-35B-A3B.
Downloads · 30 days
0
Access
Public
Updated Aug 9, 2026
Repo size
907 MB
Likes
0
Public
Click a slice to open those files.
.pt269 MB · 78%
From the Hugging Face model README
This repository contains the sparse autoencoders (SAEs), feature labels, and
direction-aware feature-ablation results produced by Kiji Inspector for
Qwen/Qwen3.6-35B-A3B.
The primary result is a layerwise causal validation of labeled SAE features.
The targeted intervention is performed in SAE latent space: for residual
activation x, the experiment computes z = encode(x), zeros the selected
coordinates to obtain z_-S, and supplies decode(z_-S) to the model. The
reconstruction-only arm supplies decode(z) without zeroing any coordinates.
The reported CATE is the paired change in contrast-tool probability:
(targeted ablation − unmodified) − (reconstruction only − unmodified).
Equivalently, for contrast-tool probability p, this is
p(decode(z_-S)) − p(decode(z)). The prompt-specific SAE round-trip effect
therefore cancels algebraically; the estimand is the incremental effect of
zeroing the selected coordinates within the SAE-reconstructed representation.
It does not claim to remove every consequence of evaluating the intervention
around decode(z) rather than the original residual x. The random arm is
not part of the primary estimand, so no assumption of uniform reconstruction
error across targeted and random features is required. CATE against the
type-conditional random arm is reported separately as a secondary specificity
control.
Produced with dataiku/kiji-inspector.
| Layer | Evaluated types | Baseline-correct prompts | Targeted flip rate | Random-control flip rate | Reconstruction-only flip rate | Direction-matched flip rate | Paired CATE (95% CI) | BH-significant types |
|---|---|---|---|---|---|---|---|---|
| 12 | 34 | 1,948 | 1.23% | 1.03% | 1.03% | 0.26% | +0.03 pp [+0.00, +0.06] | 1 / 29 |
| 20 | 35 | 1,971 | 2.38% | 2.08% | 1.88% | 0.81% | +0.02 pp [-0.02, +0.07] | 15 / 30 |
| 28 | 36 | 2,079 | 5.48% | 2.07% | 2.12% | 1.73% | +0.58 pp [+0.51, +0.65] | 25 / 31 |
| 36 | 34 | 1,925 | 13.14% | 2.65% | 2.65% | 5.45% | +2.77 pp [+2.59, +2.95] | 28 / 29 |
The table shows flip-rate point estimates; exact Clopper–Pearson confidence
intervals are stored in each ablation_report.json. Pooled CATE confidence
intervals use 2,000 paired bootstrap percentile resamples over baseline-correct
prompts, while per-type CATE intervals in the reports use 5,000 resamples.
“BH-significant” counts are based on type-level one-sided Wilcoxon signed-rank
tests of the paired targeted-minus-reconstruction probability changes, with
Benjamini–Hochberg correction within each layer. Types with fewer than 10
nonzero paired changes are not tested.
The causal signal is concentrated in the later layers. Layer 12 has no
aggregate evidence of a targeted-ablation effect (Fisher-combined
p = 0.582). Layer 20 is heterogeneous: many types have reliable positive or
negative effects, but they largely cancel in the pooled estimate. Layers 28
and 36 show clear positive aggregate CATEs, with layer 36 producing the
strongest separation.
For the 32 behavior types shared across all evaluated layers, paired CATE is +0.03 pp at layer 12, +0.05 pp at layer 20, +0.46 pp at layer 28, and +2.80 pp at layer 36. This common-type view preserves the same late-layer pattern.
Layer 20’s near-zero pooled CATE reflects genuine heterogeneity rather than an
absence of causal structure. Examples include positive effects for
urgent_vs_routine (+1.32 pp, 95% CI [+0.75, +1.92]) and
refund_vs_troubleshoot (+1.07 pp, [+0.42, +1.94]), alongside negative
effects for root_cause_vs_symptom_fix (-1.68 pp, [-2.45, -0.97]) and
spot_buy_vs_contract (-0.74 pp).
At layer 36, the largest positive effects include fundamental_vs_technical
(+11.90 pp, 95% CI [+11.10, +12.67]), spot_buy_vs_contract (+9.58 pp,
[+8.99, +10.21]), and refund_vs_troubleshoot (+9.08 pp).
manual_vs_automated_inspection is a reliable reverse-direction case
(-2.69 pp, [-3.30, -2.10]). These signed results are why feature selection and
evaluation preserve the direction of each contrast.
The random control is sampled within each behavior type, so it controls for the type-specific candidate-feature pool rather than the overall corpus rate. The targeted arm differs from reconstruction alone on 92.8–96.8% of evaluated prompts, while the matched random arm differs on 19.1–38.2%. Direction-matched flips account for 31.6% of targeted flips at layer 28 and 41.5% at layer 36.
Requested behavior types that could not be evaluated are recorded in each
layer’s ablation_report.json. They are omitted only when no feature survived
the requested direction filter:
billing_vs_technical, domestic_vs_international,
root_cause_vs_symptom_fixshort_term_vs_long_term, domestic_vs_internationaldomestic_vs_internationalshort_term_vs_long_term, domestic_vs_international,
specific_vs_broad| Item | Value |
|---|---|
| Subject model | Qwen/Qwen3.6-35B-A3B |
| Architecture | 40 layers, hidden size 2,048, hybrid linear-attention MoE, 256 experts / 8 active |
| Inspected layers | 12, 20, 28, 36 (30%, 50%, 70%, 90% depth) |
| Module family | linear_attention |
| Decision position | Prefill representation at I'll use the |
| Thinking mode | Disabled |
| Contrastive corpus | 1,574,523 pairs; 3,149,046 individual prompts; 37 behavior types |
| SAE | JumpReLU, expansion factor 4 (d_sae = 8,192), 10 epochs, target L0 = 75 |
| Activation normalization | Centered and RMS-scaled |
| Feature labeler | Qwen/Qwen3.6-35B-A3B |
| Feature judge | google/gemma-4-31b-it |
| Extraction backend | vLLM |
| Ablation backend | Hugging Face Transformers |
| Code revision | 65ae86b |
| Layer | Alive features | Dead features | Ultra-sparse features | Moderate features | Median L0 | Mean L0 | Reconstruction loss |
|---|---|---|---|---|---|---|---|
| 12 | 1,573 | 22 | 3,873 | 2,724 | 45.1 | 83.2 | 0.288 |
| 20 | 1,759 | 13 | 4,091 | 2,329 | 90.6 | 104.5 | 0.294 |
| 28 | 1,853 | 11 | 4,075 | 2,253 | 76.5 | 94.6 | 0.296 |
| 36 | 1,976 | 27 | 3,425 | 2,764 | 53.0 | 127.3 | 0.196 |
The optimization objective reached the configured L1 ceiling, so the observed L0 values should be treated as empirical SAE density rather than exact matches to the target.
| Layer | Contrastive features | Labeled features | Labeled examples | Fuzz score (95% CI) | Token score | Excellent labels |
|---|---|---|---|---|---|---|
| 12 | 849 | 1,535 | 16,885 | 0.908 [0.900, 0.915] | 0.913 | 82.3% |
| 20 | 1,185 | 1,767 | 19,437 | 0.918 [0.912, 0.924] | 0.915 | 84.5% |
| 28 | 1,149 | 1,867 | 20,537 | 0.909 [0.903, 0.916] | 0.908 | 82.7% |
| 36 | 1,331 | 2,032 | 22,352 | 0.927 [0.922, 0.932] | 0.925 | 86.3% |
| Path | Contents |
|---|---|
layer_N/sae_checkpoints/sae_final.pt | Trained JumpReLU SAE |
layer_N/sae_checkpoints/firing_rates.npy | Per-feature corpus firing rates |
layer_N/sae_checkpoints/feature_health.json | SAE health and reconstruction metrics |
layer_N/sae_checkpoints/metrics.jsonl | Per-step training metrics |
layer_N/sae_checkpoints/config.json | SAE config, mean_vec, and rms_scale |
layer_N/activations/contrastive_features.json | Ranked discriminative features and signed activation means |
layer_N/activations/feature_descriptions.json | Feature labels and descriptions |
layer_N/activations/decision_report.json | Per-contrast-type interpretation report |
layer_N/activations/fuzzing_summary.json / fuzzing_results.json | Independent label validation |
layer_N/activations/metadata.json | Latest capture metadata |
layer_N/ablation/ablation_report.json | Corrected paired causal results and raw paired arrays |
The ablation_report.json files contain pooled and per-type estimates,
confidence intervals, direction-filter diagnostics, and control outcomes. The
activations/metadata.json files document the extraction configuration used by
the corrected run.
From Kiji Inspector commit 65ae86b, point the runner at a local snapshot of
this repository and the original contrastive-pair shards:
OUTPUT_DIR=/path/to/kiji-inspector-Qwen-Qwen3.6-35B-A3B \
PAIRS_DIR=/path/to/pairs \
RUN_STEP1=auto \
bash samples/run_step1_and_ablation_qwen36.sh
RUN_STEP1=auto reuses compatible extracted activations when they are present
and regenerates them when they are absent or incompatible. The SAE and label
artifacts are reused; this command reruns the causal-ablation stage and writes
the corrected reports.
SAE.from_pretrained() downloads the checkpoint and restores its centering
and scaling parameters. encode() and decode() operate in normalized SAE
space, so callers must still use normalize_input() before encoding and
denormalize_output() after decoding:
from kiji_inspector import SAE
sae, descriptions = SAE.from_pretrained(
repo_id="575-lab/kiji-inspector-Qwen-Qwen3.6-35B-A3B",
layer=36,
)
features = sae.encode(sae.normalize_input(x))
reconstruction = sae.denormalize_output(sae.decode(features))
decode(z). With a nonlinear
downstream model, the same latent removal could have a different effect
around the original activation x. The result should therefore be read as
an SAE-defined intervention effect. A stricter follow-up can preserve the
original reconstruction residual by testing
x + [decode(z_-S) - decode(z)].The 2026-08-09 revision publishes the corrected direction-aware paired
ablation reports and refreshed extraction metadata. The SAE checkpoints and
feature labels are unchanged. Earlier ablation reports are superseded by the
ablation_report.json files currently stored in each layer directory.
If you use these artifacts, cite Kiji Inspector and this model repository.