---
license: other
base_model: google/gemma-4-26B-A4B-it
base_model_relation: adapter
tags:
- metacognition
- error-detection
- adapter
- aether
- vidraft-darwin-chimera
---

# AETHER Metacognition Adapter — gemma-4-26B-A4B-it

A lightweight **metacognition *adapter*** (this is an adapter, **not** a fine-tune) for [`google/gemma-4-26B-A4B-it`](https://huggingface.co/google/gemma-4-26B-A4B-it).
The base model's weights stay **frozen and unchanged** — the adapter only reads the base model's internal state to predict **when the base model is about to make a mistake**.

## Platform & technology
Produced on **VIDRAFT's Darwin / Chimera model-generation platform**, with VIDRAFT's proprietary **AETHER metacognition-emergence technology** grafted on. The adapter surfaces a calibrated *"am I about to be wrong?"* signal that the base model's own confidence does not provide.

## Why it matters
On free-form tasks a model's own confidence is a weak signal of correctness. This adapter recovers that signal, so a system can defer, double-check, or escalate exactly when the model is likely wrong.

## Scores — AETHER Metacognition Benchmark (free-form, held-out)
| Metric | Value |
|--------|-------|
| **Adapter gain** (Δ AUROC vs the model's own confidence) | **+0.258** |
| Error-detection AUROC (adapter) | 0.806 |
| Base-confidence AUROC | 0.548 |

A positive gain means this adapter detects the model's errors **better than the model's own confidence.**

## Usage
```python
import torch, torch.nn as nn
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "google/gemma-4-26B-A4B-it"
REPO = "FINAL-Bench/metacog-adapter-gemma-4-26B-A4B-it"

tok = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(BASE, dtype="auto", device_map="auto").eval()

# Metacognition adapter = base model's last hidden state -> P(this answer is wrong). Base stays frozen.
d = model.config.hidden_size
adapter = nn.Sequential(nn.LayerNorm(d), nn.Linear(d, d // 4), nn.GELU(), nn.Dropout(0.1), nn.Linear(d // 4, 1))
adapter.load_state_dict(load_file(hf_hub_download(REPO, "adapter.safetensors")))
adapter.eval().to(model.device, dtype=torch.float32)

prompt = "..."  # your question / task
ids = tok.apply_chat_template([{"role": "user", "content": prompt}],
                              return_tensors="pt", add_generation_prompt=True).to(model.device)
with torch.no_grad():
    h = model(ids, output_hidden_states=True).hidden_states[-1][0, -1].float()
    p_wrong = torch.sigmoid(adapter(h)).item()
print(f"P(model is about to be wrong) = {p_wrong:.3f}")   # higher => defer / double-check / escalate
```

## Files
- `adapter.safetensors` — adapter weights (base model frozen)
- `config.json` — metadata (base model, hidden size)

## Links
- 🏆 [Metacognition Leaderboard](https://huggingface.co/spaces/ginigen-ai/Metacognition-Leaderboard-Space)
- 📚 Collection: **AETHER Metacognition Adapters** (FINAL-Bench)

---
*Proprietary — "AETHER Metacognition Adapter (proprietary)". Weights released for evaluation; the training/inference method (VIDRAFT Darwin/Chimera platform + AETHER metacognition emergence) is proprietary and not included.*