Mechanistic Interpretability Nnsight 教程

# nnsight 教程 ## 教程 1：基础激活分析 ### 目标 加载模型，访问内部激活值并进行分析。 ### 步骤 python from nnsight import LanguageModel import torch # 1. 加载模型 model = LanguageModel("openai-community/gpt2", device_map="auto") # 2. 追踪并收集激活值 prompt = "The capital of France is" with model.trace(prompt) as tracer: # 从多个层收集 activations = {} for i in range(12): # GPT-2 有 12 层 activations[i] = model.transformer.h[i].output[0].save() # 获取最终 logits logits = model.output.save() # 3. 分析（在上下文之外） print("逐层激活范数:") for layer, act in activations.items(): print(f" Layer {layer}: {act.norm().item():.2f}") # 4. 检查预测结果 probs = torch.softmax(logits[0, -1], dim=-1) top_tokens = probs.topk(5) print("nTop 预测:") for token_id, prob in zip(top_tokens.indices, top_tokens.values): token_str = model.tokenizer.decode(token_id) print(f" {token_str!r}: {prob.item():.3f}") --- ## 教程 2：激活修补 (Activation Patching) ### 目标 将一个提示词的激活值修补到另一个提示词中，以测试因果关系。 ### 步骤 python from nnsight import LanguageModel import torch model = LanguageModel("gpt2", device_map="auto") clean_prompt = "The Eiffel Tower is in the city of" corrupted_prompt = "The Colosseum is in the city of" # 1. 获取干净的激活值 with model.trace(clean_prompt) as tracer: clean_hidden = model.transformer.h[8].output[0].save() clean_logits = model.output.save() # 2. 定义度量指标 paris_token = model.tokenizer.encode(" Paris")[0] rome_token = model.tokenizer.encode(" Rome")[0] def logit_diff(logits): return (logits[0, -1, paris_token] - logits[0, -1, rome_token]).item() print(f"干净的 logit 差值: {logit_diff(clean_logits):.3f}") # 3. 将干净的激活值修补到损坏的提示词中 with model.trace(corrupted_prompt) as tracer: # 用干净的激活值替换第 8 层输出 model.transformer.h[8].output[0][:] = clean_hidden patched_logits = model.output.save() print(f"修补后的 logit 差值: {logit_diff(patched_logits):.3f}") # 4. 系统性修补扫描 results = torch.zeros(12) # 12 层 for layer in range(12): # 获取该层的干净激活值 with model.trace(clean_prompt) as tracer: clean_act = model.transformer.h[layer].output[0].save() # 修补到损坏的提示词中 with model.trace(corrupted_prompt) as tra