跳到主要内容

稀疏自编码器训练

提供使用 SAELens 训练和分析稀疏自编码器(SAE)的指南,用于将神经网络激活分解为可解释的特征。适用于在语言模型中发现可解释特征、分析超位置(superposition)或研究单义性(monosemantic)表示时。

技能元数据

来源可选 — 使用 hermes skills install official/mlops/saelens 安装
路径optional-skills/mlops/saelens
版本1.0.0
作者Orchestra Research
许可证MIT
依赖项sae-lens>=6.0.0, transformer-lens>=2.0.0, torch>=2.0.0
标签Sparse Autoencoders, SAE, Mechanistic Interpretability, Feature Discovery, Superposition

参考:完整 SKILL.md

信息

以下是 Hermes 在触发此技能时加载的完整技能定义。这是技能激活时代理看到的指令。

SAELens:用于机械可解释性的稀疏自编码器

SAELens 是用于训练和分析稀疏自编码器(SAE)的主要库——这是一种将多义性(polysemantic)神经网络激活分解为稀疏、可解释特征的技术。基于 Anthropic 在单义性方面的突破性研究。

GitHub: jbloomAus/SAELens (1,100+ stars)

问题:多义性与超位置

神经网络中的单个神经元是多义的——它们在多个语义不同的上下文中激活。这是因为模型使用超位置来表示比其神经元数量更多的特征,使得可解释性变得困难。

SAE 通过以下方式解决此问题:将密集激活分解为稀疏的单义特征——通常对于任何给定输入,只有少量特征被激活,且每个特征对应一个可解释的概念。

何时使用 SAELens

在需要执行以下操作时使用 SAELens:

  • 发现模型激活中的可解释特征
  • 理解模型学到了哪些概念
  • 研究超位置和特征几何
  • 执行基于特征的引导(steering)或消融
  • 分析与安全相关的特征(欺骗、偏见、有害内容)

在以下情况考虑替代方案:

  • 你需要基本的激活分析 → 直接使用 TransformerLens
  • 你想要因果干预实验 → 使用 pyveneTransformerLens
  • 你需要生产环境引导 → 考虑直接激活工程

安装

pip install sae-lens

要求:Python 3.10+, transformer-lens>=2.0.0

核心概念

SAE 学习的内容

SAE 经过训练,通过稀疏瓶颈重建模型激活:

Input Activation → Encoder → Sparse Features → Decoder → Reconstructed Activation
(d_model) ↓ (d_sae >> d_model) ↓ (d_model)
sparsity reconstruction
penalty loss

损失函数MSE(original, reconstructed) + L1_coefficient × L1(features)

关键验证(Anthropic 研究)

在《Towards Monosemanticity》中,人类评估者发现 70% 的 SAE 特征具有真正的可解释性。发现的特征包括:

  • DNA 序列、法律语言、HTTP 请求
  • 希伯来语文本、营养说明、代码语法
  • 情感、命名实体、语法结构

工作流 1:加载和分析预训练 SAE

分步指南

from transformer_lens import HookedTransformer
from sae_lens import SAE

# 1. Load model and pre-trained SAE
model = HookedTransformer.from_pretrained("gpt2-small", device="cuda")
sae, cfg_dict, sparsity = SAE.from_pretrained(
release="gpt2-small-res-jb",
sae_id="blocks.8.hook_resid_pre",
device="cuda"
)

# 2. Get model activations
tokens = model.to_tokens("The capital of France is Paris")
_, cache = model.run_with_cache(tokens)
activations = cache["resid_pre", 8] # [batch, pos, d_model]

# 3. Encode to SAE features
sae_features = sae.encode(activations) # [batch, pos, d_sae]
print(f"Active features: {(sae_features > 0).sum()}")

# 4. Find top features for each position
for pos in range(tokens.shape[1]):
top_features = sae_features[0, pos].topk(5)
token = model.to_str_tokens(tokens[0, pos:pos+1])[0]
print(f"Token '{token}': features {top_features.indices.tolist()}")

# 5. Reconstruct activations
reconstructed = sae.decode(sae_features)
reconstruction_error = (activations - reconstructed).norm()

可用的预训练 SAE

发布版本模型
gpt2-small-res-jbGPT-2 Small多个残差流
gemma-2b-resGemma 2B残差流
HuggingFace 上的各种版本搜索标签 saelens各种

检查清单

  • 使用 TransformerLens 加载模型
  • 为目标层加载匹配的 SAE
  • 将激活编码为稀疏特征
  • 识别每个 token 激活最高的特征
  • 验证重建质量

工作流 2:训练自定义 SAE

分步指南

from sae_lens import SAE, LanguageModelSAERunnerConfig, SAETrainingRunner

# 1. Configure training
cfg = LanguageModelSAERunnerConfig(
# Model
model_name="gpt2-small",
hook_name="blocks.8.hook_resid_pre",
hook_layer=8,
d_in=768, # Model dimension

# SAE architecture
architecture="standard", # or "gated", "topk"
d_sae=768 * 8, # Expansion factor of 8
activation_fn="relu",

# Training
lr=4e-4,
l1_coefficient=8e-5, # Sparsity penalty
l1_warm_up_steps=1000,
train_batch_size_tokens=4096,
training_tokens=100_000_000,

# Data
dataset_path="monology/pile-uncopyrighted",
context_size=128,

# Logging
log_to_wandb=True,
wandb_project="sae-training",

# Checkpointing
checkpoint_path="checkpoints",
n_checkpoints=5,
)

# 2. Train
trainer = SAETrainingRunner(cfg)
sae = trainer.run()

# 3. Evaluate
print(f"L0 (avg active features): {trainer.metrics['l0']}")
print(f"CE Loss Recovered: {trainer.metrics['ce_loss_score']}")

关键超参数

参数典型值效果
d_sae4-16× d_model更多特征,更高容量
l1_coefficient5e-5 到 1e-4越高 = 越稀疏,准确度越低
lr1e-4 到 1e-3标准优化器学习率
l1_warm_up_steps500-2000防止早期特征死亡

评估指标

指标目标含义
L050-200每个 token 的平均激活特征数
CE Loss Score80-95%相对于原始值的交叉熵恢复率
Dead Features<5%从未激活的特征
Explained Variance>90%重建质量

检查清单

  • 选择目标层和钩子点(hook point)
  • 设置扩展因子(d_sae = 4-16× d_model)
  • 调整 L1 系数以获得所需的稀疏度
  • 启用 L1 预热以防止特征死亡
  • 在训练期间监控指标(W&B)
  • 验证 L0 和 CE 损失恢复
  • 检查死亡特征比例

工作流 3:特征分析和引导

分析单个特征

from transformer_lens import HookedTransformer
from sae_lens import SAE
import torch

model = HookedTransformer.from_pretrained("gpt2-small", device="cuda")
sae, _, _ = SAE.from_pretrained(
release="gpt2-small-res-jb",
sae_id="blocks.8.hook_resid_pre",
device="cuda"
)

# Find what activates a specific feature
feature_idx = 1234
test_texts = [
"The scientist conducted an experiment",
"I love chocolate cake",
"The code compiles successfully",
"Paris is beautiful in spring",
]

for text in test_texts:
tokens = model.to_tokens(text)
_, cache = model.run_with_cache(tokens)
features = sae.encode(cache["resid_pre", 8])
activation = features[0, :, feature_idx].max().item()
print(f"{activation:.3f}: {text}")

特征引导

def steer_with_feature(model, sae, prompt, feature_idx, strength=5.0):
"""Add SAE feature direction to residual stream."""
tokens = model.to_tokens(prompt)

# Get feature direction from decoder
feature_direction = sae.W_dec[feature_idx] # [d_model]

def steering_hook(activation, hook):
# Add scaled feature direction at all positions
activation += strength * feature_direction
return activation

# Generate with steering
output = model.generate(
tokens,
max_new_tokens=50,
fwd_hooks=[("blocks.8.hook_resid_pre", steering_hook)]
)
return model.to_string(output[0])

特征归因

# Which features most affect a specific output?
tokens = model.to_tokens("The capital of France is")
_, cache = model.run_with_cache(tokens)

# Get features at final position
features = sae.encode(cache["resid_pre", 8])[0, -1] # [d_sae]

# Get logit attribution per feature
# Feature contribution = feature_activation × decoder_weight × unembedding
W_dec = sae.W_dec # [d_sae, d_model]
W_U = model.W_U # [d_model, vocab]

# Contribution to "Paris" logit
paris_token = model.to_single_token(" Paris")
feature_contributions = features * (W_dec @ W_U[:, paris_token])

top_features = feature_contributions.topk(10)
print("Top features for 'Paris' prediction:")
for idx, val in zip(top_features.indices, top_features.values):
print(f" Feature {idx.item()}: {val.item():.3f}")

常见问题与解决方案

问题:高死特征比例

# WRONG: No warm-up, features die early
cfg = LanguageModelSAERunnerConfig(
l1_coefficient=1e-4,
l1_warm_up_steps=0, # Bad!
)

# RIGHT: Warm-up L1 penalty
cfg = LanguageModelSAERunnerConfig(
l1_coefficient=8e-5,
l1_warm_up_steps=1000, # Gradually increase
use_ghost_grads=True, # Revive dead features
)

问题:重建效果差(交叉熵恢复率低)

# Reduce sparsity penalty
cfg = LanguageModelSAERunnerConfig(
l1_coefficient=5e-5, # Lower = better reconstruction
d_sae=768 * 16, # More capacity
)

问题:特征不可解释

# Increase sparsity (higher L1)
cfg = LanguageModelSAERunnerConfig(
l1_coefficient=1e-4, # Higher = sparser, more interpretable
)
# Or use TopK architecture
cfg = LanguageModelSAERunnerConfig(
architecture="topk",
activation_fn_kwargs={"k": 50}, # Exactly 50 active features
)

问题:训练期间出现内存错误

cfg = LanguageModelSAERunnerConfig(
train_batch_size_tokens=2048, # Reduce batch size
store_batch_size_prompts=4, # Fewer prompts in buffer
n_batches_in_buffer=8, # Smaller activation buffer
)

与 Neuronpedia 集成

neuronpedia.org 浏览预训练的 SAE 特征:

# Features are indexed by SAE ID
# Example: gpt2-small layer 8 feature 1234
# → neuronpedia.org/gpt2-small/8-res-jb/1234

关键类参考

用途
SAE稀疏自编码器模型
LanguageModelSAERunnerConfig训练配置
SAETrainingRunner训练循环管理器
ActivationsStore激活值收集与批处理
HookedSAETransformerTransformerLens + SAE 集成

参考文档

有关详细的 API 文档、教程和高级用法,请参阅 references/ 文件夹:

文件内容
references/README.md概述和快速入门指南
references/api.mdSAE、TrainingSAE、配置的完整 API 参考
references/tutorials.md训练、分析、 steering 的分步教程

外部资源

教程

论文

官方文档

SAE 架构

架构描述用例
StandardReLU + L1 惩罚通用目的
Gated学习门控机制更好的稀疏性控制
TopK恰好 K 个活跃特征一致的稀疏性
# TopK SAE (exactly 50 features active)
cfg = LanguageModelSAERunnerConfig(
architecture="topk",
activation_fn="topk",
activation_fn_kwargs={"k": 50},
)