You can deploy this model and start earning money today!

ceselder/
qwen3-8b-sae-maxact-lora

Name: qwen3-8b-sae-maxact-lora
Availability: InStock

max-activating-examples

Views

No views yet

Qwen3-8B SAE max-activating-example generator (LoRA)

Generates text that maximally activates a target SAE feature of Qwen3-8B itself, conditioned activation-oracle-style: the target direction is injected into the layer-1 residual stream, and the model writes a snippet that drives that direction at layer 27.

Base: Qwen/Qwen3-8B (rsLoRA r=64, α=16, all-linear)
SAE: adamkarvonen/qwen3-8b-saes — BatchTopK, resid_post_layer_27/trainer_2 (65,536 features)
Conditioning vector: the unit-normalized SAE encoder column W_enc[:, f] of the target feature
Training: activation-oracle init → SFT on 211k examples (corpus top-3 per feature + 21.4k best-of-64 self-distilled winners) → 10k GRPO steps (group-relative PG, reward = max-over-positions centered dot with the target direction, measured on the standalone re-tokenized completion through the clean base model; fluency/distinct gates, length penalty, clipped KL to init)

Protocol

Prompt (chat template, enable_thinking=False), with a single placeholder token ?:

Layer: 27
 ?
The token above contains an injected direction from this model's residual stream at the indicated layer. Write one short text snippet (roughly 30 tokens or fewer) that would make this direction activate as strongly as possible. Output only the snippet itself, with no explanation, preamble, or quotation marks.

At the placeholder position, add the norm-matched direction into the layer-1 residual stream during prefill:

resid[b, pos] += v_hat * resid[b, pos].norm() * 1.0 # v_hat = unit encoder column

Sample with temperature 1.0 (best-of-N over 8–64 samples strongly recommended; scoring = re-tokenize the completion standalone, run the clean base model, take max-over-positions activation).

Results (500–1000 held-out features, fair truncation-aware corpus references)

metric	value
greedy: mean activation vs corpus-best window (500M tokens searched)	~0.61
greedy: fraction of features beating corpus best	~15%
best-of-16 @ T=1: fraction beating corpus best	~29%
best-of-64 (earlier checkpoint lineage)	~39%

Trained/evaluated on disjoint feature splits. This checkpoint (GRPO step 10,000) maximizes held-out best-of-N performance; later checkpoints improve train reward but not transfer.

Part of a MATS interpretability project (Neel Nanda stream) on eliciting model self-knowledge of its own feature space.

ceselder/qwen3-8b-sae-maxact-lora

Qwen3-8B SAE max-activating-example generator (LoRA)

Protocol

Results (500–1000 held-out features, fair truncation-aware corpus references)

ceselder/
qwen3-8b-sae-maxact-lora