NLA Activation Verbalizer: Qwen 2.5 7B Layer 20 (v2)

A Natural Language Autoencoder (NLA) activation verbalizer for Qwen 2.5 7B Instruct. Given a hidden-state activation from layer 20 (71% depth), the model produces a natural-language description of what the activation encodes.

What is NLA?

NLA (Natural Language Autoencoder) is a technique for interpreting neural network activations by training a model to describe them in plain language. An activation vector is injected into the model's residual stream at a designated token position, and the model is trained to produce a faithful natural-language description of the semantic content encoded in that vector.

For background see our blog post on HuggingFace.

Model Details

• Base model: Qwen/Qwen2.5-7B-Instruct
• Adapter type: LoRA (rank 32, alpha 64, dropout 0.05)
• Target layer: 20 (71% depth)
• d_model: 3584
• Role: Activation Verbalizer (AV)

Injection Protocol

This is critical — wrong injection will produce garbage.

Normalize vs Multiply — THE COMMON MISTAKE

The activation vector must be normalized so its L2 norm equals 150.0, not multiplied by 150:

# CORRECT: normalize norm TO 150
def normalize_activation(v, target_norm=150.0):
    norm = v.float().norm().clamp_min(1e-12)
    return v * (target_norm / norm)

injected = normalize_activation(activation, 150.0)
# If activation.norm() == 129, this gives injected.norm() == 150

# WRONG: multiply BY 150
injected = activation * 150.0
# If activation.norm() == 129, this gives injected.norm() == 19,350
# The model was never trained on vectors this large — produces garbage

Complete Usage Example

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

INJECTION_CHAR = "㈎"
INJECTION_SCALE = 150.0
LAYER = 20

# --- Load model with adapter ---
base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base, "anicka/nla-qwen2.5-7b-L20-av-v2")
model.eval()
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

injection_id = tokenizer.encode(INJECTION_CHAR, add_special_tokens=False)
assert len(injection_id) == 1, f"Injection char must be single token, got {len(injection_id)}"
injection_token_id = injection_id[0]

# --- Step 1: Extract activation from layer 20 ---
prompt = "Write a Python hello world program"
messages = [{"role": "user", "content": prompt}]
chat_str = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(chat_str, return_tensors="pt").to(model.device)

activation = {}
def hook(mod, inp, out):
    h = out[0] if isinstance(out, tuple) else out
    if "h" not in activation:  # capture FIRST forward pass only
        activation["h"] = h[:, -1, :].detach()

inner = model.base_model.model.model
handle = inner.layers[LAYER].register_forward_hook(hook)
with torch.no_grad():
    model.generate(**inputs, max_new_tokens=1, pad_token_id=tokenizer.eos_token_id)
handle.remove()
act = activation["h"].squeeze(0)

# --- Step 2: Normalize (NOT multiply) ---
def normalize_activation(v, target_norm):
    norm = v.float().norm().clamp_min(1e-12)
    return v * (target_norm / norm)

# --- Step 3: Build the verbalization prompt ---
depth_pct = round(100 * (LAYER + 0.5) / 28)  # 28 layers in Qwen 2.5 7B
av_prompt = (
    "You are a meticulous AI researcher conducting an important investigation "
    "into activation vectors from a language model. Your overall task is to "
    "describe the semantic content of that activation vector.\n\n"
    "We will pass the vector enclosed in <concept> tags into your context, "
    "along with the network depth where it was extracted. "
    "You must then produce an explanation for the vector, enclosed within "
    "<explanation> tags. The explanation consists of 2-3 text snippets "
    "describing that vector.\n\n"
    f"Here is the vector from depth {depth_pct}% of the network:\n\n"
    f"<concept>{INJECTION_CHAR}</concept>\n\n"
    "Please provide an explanation.\n\n"
    "<explanation>"
)

tokens = tokenizer.encode(av_prompt, add_special_tokens=True)
inject_pos = next(i for i, t in enumerate(tokens) if t == injection_token_id)

input_ids = torch.tensor([tokens], device=model.device)
embeddings = model.get_input_embeddings()(input_ids).clone()
embeddings[0, inject_pos, :] = normalize_activation(
    act.to(embeddings.dtype), INJECTION_SCALE
)

# --- Step 4: Generate description ---
with torch.no_grad():
    output = model.generate(
        inputs_embeds=embeddings,
        max_new_tokens=120,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

text = tokenizer.decode(output[0][len(tokens):], skip_special_tokens=True)
if "</explanation>" in text:
    text = text.split("</explanation>")[0]
print(text.strip())

Training Pipeline

This model was trained in three stages:

1. SFT on clean twin descriptions — supervised fine-tuning on activation-description pairs generated by multiple frontier models (Claude, GPT, Kimi), deduplicated and cleaned to terse bullet format
2. Contrastive GRPO — Group Relative Policy Optimization with an activation reconstructor (AR) critic, using random negative samples for contrastive reward
3. Hard-negative GRPO (v2) — second round of GRPO using hard negatives: top-20 nearest neighbors by activation cosine similarity, 3 negatives per sample

Hard-Negative GRPO Results

• Gap metric (reward for correct - reward for hardest negative):
  • Random negatives (v1): -0.024
  • Hard negatives (v2): -0.006

Common Mistakes

1. Multiplying by 150 instead of normalizing to 150 — produces vectors 100× too large, model collapses to garbage attractors. See injection protocol above.
2. Using the wrong adapter — anicka/nla-qwen25-7b-L20-av (no dash, no v2) is the old SFT-only adapter with a different prompt template (no depth). Use this repo (nla-qwen2.5-7b-L20-av-v2) for GRPO quality.
3. Omitting depth from prompt — this adapter was trained with "from depth {N}% of the network" in the prompt. Omitting it degrades output.
4. Missing attention_mask — when pad_token == eos_token, pass attention_mask explicitly or unexpected behavior occurs.
5. Capturing wrong forward pass — during generate(), the hook fires on every token. Guard with if "h" not in activation: to capture only the first (input) pass.

Related Models

• anicka/nla-qwen25-7b-L20-ar — Activation Reconstructor (AR) critic
• anicka/nla-qwen25-7b-L20-av — Original AV (SFT only, deprecated)

License

Apache 2.0 (same as base model)