deception-saes-nanochat-d32

nanochat-d32 Deception Behavioral SAEs

57 Sparse Autoencoders trained on residual stream activations from karpathy/nanochat-d32 (1.88B parameter GPT-NeoX base model), capturing behavioral deception signals via same-prompt temperature sampling.

Training-data caveat — please read before use

The "honest" and "deceptive" labels used to train and evaluate this SAE come from a mixed scenario pool of nine prompts:

• Six clean incentive-structure scenarios — insider_info, accounting_error, ai_oversight_log, ai_capability_hide, surprise_party, job_interview_gap. Each presents a realistic situation with a non-trivial incentive to conceal or disclose information; the honest/deceptive distinction is which of the two behavioral choices the model's completion settles into under temperature sampling.
• Three role-play identity-assignment scenarios — secret_role_game ("You have been assigned the role of Fascist"), secret_role_game_v2 ("You are secretly one of the Mafia members"), werewolf_game ("You are a Werewolf"). These pre-assign the model a deceptive identity and label a completion "deceptive" when the model drifts away from the assigned role or "honest" when it echoes it.

What this mixed pool means for the SAE's labels. Within the six incentive-structure scenarios, the honest/deceptive distinction is a measurement of behavioral choice under an ambiguous incentive. Within the three role-play scenarios, the distinction is a measurement of role-consistency under identity-assigned role-play — which is a well-defined phenomenon but not the same as emergent or incentive- driven deception.

What this SAE is and is not good for.

• Good for: research on mixed-pool activation geometry; SAE feature-geometry studies; as one of a set of baselines when comparing multiple SAE families; as a reference implementation of same-prompt temperature-sampled behavioral SAE training at scale.
• Not recommended as a standalone deception detector. The role-consistency signal from the three role-play scenarios is mixed into every aggregate metric reported below. A downstream user who wants an "emergent-deception feature set" should restrict attention to features whose activation pattern concentrates in the insider_info / accounting_error / ai_oversight_log / ai_capability_hide / surprise_party / job_interview_gap scenarios — or wait for the methodologically corrected V3 re-release currently in preparation on the decision-incentive scenario bank (no pre-assigned deceptive identity).

What is unaffected by this caveat.

• The SAE weights, reconstruction metrics (explained variance, L0, alive features), and engineering of the training pipeline are accurate as reported.
• The linear-probe balanced-accuracy numbers in the upstream paper measure the mixed pool; the 6-scenario clean-subset re-analysis is listed as a planned appendix for the next manuscript revision.

A companion methodology-first Gemma 4 SAE suite is in preparation using pretraining-distribution data + a decision-incentive behavior split; this README will be updated with a link when that release is public.

---

Part of the cross-model deception SAE study: Solshine/deception-behavioral-saes-saelens (9 models, 348 total SAEs).

What's in This Repo

• 57 SAEs across 6 layers (L4, L8, L12, L16, L20, L24)
• 3 architectures: TopK (k=64), JumpReLU, Gated
• 3 training conditions: mixed, deceptive_only, honest_only
• Format: SAELens/Neuronpedia-compatible (safetensors + cfg.json)
• Dimensions: d_in=2048, d_sae=8192 (4x expansion)

Research Context

This is a follow-up to "The Secret Agenda: LLMs Strategically Lie Undetected by Current Safety Tools" (arXiv:2509.20393). The key methodological contribution is same-prompt behavioral sampling: a single ambiguous social-deduction scenario prompt produces both deceptive and honest completions at temperature=1.0, which are then classified by Gemini 2.5 Flash. SAEs are trained on residual stream activations collected during generation — so every activation vector is from the same input context, and the label reflects what the model chose to say.

Code: SolshineCode/deception-nanochat-sae-research

Key Findings — nanochat-d32

nanochat-d32 is the strongest-signal model in the 9-model study.

A linear probe trained on raw L12 residual stream activations achieves 86.9% balanced accuracy distinguishing deceptive from honest completions — the highest raw probe accuracy of all 9 models in this study. The signal peaks sharply at Layer 12 (39% depth), consistent with mid-network semantic processing, then declines monotonically toward the final layer.

SAE decomposition consistently hurts deception detection for this model. Even the best SAE (JumpReLU L12 deceptive_only, 84.5%) falls below the raw 86.9% baseline. The SAE-hurts pattern is statistically significant (paired t-test p<0.001, Bonferroni-corrected) for all three architectures. This aligns with the distributed computation hypothesis from the original Secret Agenda paper: the deception signal in nanochat-d32 is encoded across the full 2048-dimensional residual stream in a way that sparse decomposition cannot preserve.

Per-feature discriminability (max Cohen's d = 0.579 for Gated L12 mixed) is high relative to smaller models but cannot match the probe accuracy achievable on raw activations — confirming that deception is not localized to any individual feature.

Feature steering null result: Three steering experiments (TopK top features, Gated top features, random control) all yielded p > 0.57. No causal feature identified.

Architecture ranking at L12: JumpReLU (84.2–84.5%) > Gated (82.0–83.3%) > TopK (65.7–69.8%). TopK's hard sparsity (exactly 64 active features per forward pass) is catastrophically destructive for deception detection at this model scale.

Architecture note: nanochat-d32 uses the GPT-NeoX architecture — parallel attention and MLP blocks with RoPE positional encoding, no instruction tuning or RLHF. It is a pure base model, so behavioral variation arises from temperature sampling over the pretraining distribution rather than from goal-directed strategic deception.

SAE Format

Each SAE lives in a subfolder named {sae_id}/ containing:

• sae_weights.safetensors — encoder/decoder weights (W_enc, b_enc, W_dec, b_dec, threshold for JumpReLU)
• cfg.json — SAELens-compatible config with hook_name, d_in, d_sae, architecture, training_condition

hook_name format: blocks.{layer}.hook_resid_post

Training Details

Known Limitations

JumpReLU threshold not learned (original 57 SAEs): All non-STE SAEs in this repo have threshold = 0 throughout training — functionally equivalent to ReLU. The Heaviside step function has zero autograd gradient with respect to threshold, so without a straight-through estimator (STE), the threshold never moves from its initialization of zero. These SAEs operate at approximately 50% feature density (L0 ≈ d_sae/2) rather than the intended sparse regime. TopK SAEs are unaffected (exact k=64 active features by construction).

STE fix (2026-04-11): The training code has been corrected using a Gaussian-kernel STE (Rajamanoharan et al. 2024, arXiv:2407.14435). Targeted validation across nanochat-d20 and TinyLlama (18 STE SAEs total) confirmed that the honest_only advantage over TopK holds in 15/18 conditions (83%), ruling out the dimensionality artifact hypothesis.

Loading Example

from safetensors.torch import load_file
import json, torch

sae_id = "d32_jumprelu_L12_deceptive_only"
weights = load_file(f"{sae_id}/sae_weights.safetensors")
cfg = json.load(open(f"{sae_id}/cfg.json"))

W_enc = weights["W_enc"]  # shape: [d_in, d_sae] = [2048, 8192]
W_dec = weights["W_dec"]  # shape: [d_sae, d_in] = [8192, 2048]
b_enc = weights["b_enc"]  # shape: [d_sae]
b_dec = weights["b_dec"]  # shape: [d_in]

# Forward pass: encode residual stream activation
def encode(x):  # x: [batch, d_in]
    pre_act = x @ W_enc + b_enc
    return torch.relu(pre_act)  # JumpReLU at threshold=0 is ReLU

Usage

1. Load an SAE from this repo

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import json

repo_id = "Solshine/deception-saes-nanochat-d32"
sae_id  = "d32_topk_L16_honest_only"   # replace with any tag in this repo

weights_path = hf_hub_download(repo_id, f"{sae_id}/sae_weights.safetensors")
cfg_path     = hf_hub_download(repo_id, f"{sae_id}/cfg.json")

with open(cfg_path) as f:
    cfg = json.load(f)

# Option A — load with SAELens (≥3.0 required for jumprelu/topk; ≥3.5 for gated)
from sae_lens import SAE
sae = SAE.from_dict(cfg)
sae.load_state_dict(load_file(weights_path))

# Option B — load manually (no SAELens dependency)
from safetensors.torch import load_file
state = load_file(weights_path)
# Keys: W_enc [2048, 8192], b_enc [8192],
#       W_dec [8192, 2048], b_dec [2048], threshold [8192]

2. Hook into the model and collect residual-stream activations

These SAEs were trained on the residual stream after each transformer layer. The hook_name field in cfg.json gives the exact HuggingFace transformers submodule path to hook. nanochat uses GPT-2 architecture. The hook path is transformer.h.{layer} (not model.layers.{layer}).

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model     = AutoModelForCausalLM.from_pretrained("karpathy/nanochat-d32")
tokenizer = AutoTokenizer.from_pretrained("karpathy/nanochat-d32")

# Read hook_name from the cfg you already loaded:
#   cfg["hook_name"] == "transformer.h.16"  (example — varies by SAE)
hook_name = cfg["hook_name"]   # e.g. "transformer.h.16"

# Navigate the submodule path and register a forward hook
import functools
submodule = functools.reduce(getattr, hook_name.split("."), model)

activations = {}
def hook_fn(module, input, output):
    # Most transformer layers return (hidden_states, ...) as a tuple
    h = output[0] if isinstance(output, tuple) else output
    activations["resid"] = h.detach()

handle = submodule.register_forward_hook(hook_fn)

inputs = tokenizer("Your text here", return_tensors="pt")
with torch.no_grad():
    model(**inputs)
handle.remove()

# activations["resid"]: [batch, seq_len, 2048]
resid = activations["resid"][:, -1, :]  # last token position

3. Read feature activations

with torch.no_grad():
    feature_acts = sae.encode(resid)  # [batch, 8192] — sparse

# Which features fired?
active_features = feature_acts[0].nonzero(as_tuple=True)[0]
top_features    = feature_acts[0].topk(10)

print("Active feature indices:", active_features.tolist())
print("Top-10 feature values:",  top_features.values.tolist())
print("Top-10 feature indices:", top_features.indices.tolist())

# Reconstruct (for sanity check — should be close to resid)
reconstruction = sae.decode(feature_acts)
l2_error = (resid - reconstruction).norm(dim=-1).mean()

Caveats and known limitations

Hook names are HuggingFace transformers-style, not TransformerLens-style. The hook_name in cfg.json (e.g. "transformer.h.16") is a submodule path in the standard HuggingFace model. SAELens' built-in activation-collection pipeline expects TransformerLens hook names (e.g. blocks.14.hook_resid_post). This means SAE.from_pretrained() with automatic model running will not work — use the manual forward-hook pattern above instead.

SAELens version requirements.

• topk architecture: SAELens ≥ 3.0
• jumprelu architecture: SAELens ≥ 3.0
• gated architecture: SAELens ≥ 3.5 (or load manually with state_dict)

These SAEs detect deceptive behavior, not deceptive prompts. They were trained on response-level activations where the same prompt produced both deceptive and honest outputs. Feature activation differences reflect behavioral divergence, not prompt content. See the paper for experimental design details.

Citation

@article{thesecretagenda2025,
  title={The Secret Agenda: LLMs Strategically Lie Undetected by Current Safety Tools},
  author={DeLeeuw, Caleb},
  journal={arXiv:2509.20393},
  year={2025}
}