---
license: gemma
language:
- en
pipeline_tag: text-generation
base_model: google/gemma-4-E2B
tags:
- error-correction
- frozen-backbone
- capability-preservation
- logit-adjustment
- parameter-efficient
- gemma
- apple-silicon
- on-device-ai
- cehri
model-index:
- name: Prajna-CRNv2
  results:
  - task:
      type: question-answering
      name: CEHRI Exam (original, 60 Qs)
    dataset:
      type: custom
      name: CEHRI-60
    metrics:
    - type: accuracy
      value: 53.3
  - task:
      type: question-answering
      name: CEHRI Exam (reworded, 120 Qs)
    dataset:
      type: custom
      name: CEHRI-120-reworded
    metrics:
    - type: accuracy
      value: 43.3
  - task:
      type: multiple-choice
      name: MMLU (sampled, N=200)
    dataset:
      type: cais/mmlu
      name: MMLU
    metrics:
    - type: accuracy
      value: 62.5
  - task:
      type: question-answering
      name: BoolQ (sampled, N=200)
    dataset:
      type: google/boolq
      name: BoolQ
    metrics:
    - type: accuracy
      value: 72.0
---

# 🪷 Prajna-CRNv2 — Safe Error Correction for Language Models

<p>
<a href="https://github.com/eulogik/prajna"><img alt="GitHub" src="https://img.shields.io/badge/GitHub-eulogik%2Fprajna-black?style=for-the-badge&logo=github"></a> <img alt="HF downloads" src="https://img.shields.io/badge/dynamic/json?url=https://huggingface.co/api/models/eulogik/Prajna-CRNv2&query=downloads&color=blue&label=HF%20downloads&style=for-the-badge"> <img alt="Correction" src="https://img.shields.io/badge/CEHRI%20correction-53.3%25-6a4cff?style=for-the-badge"> <img alt="Degradation" src="https://img.shields.io/badge/capability%20loss-0%25-00d26a?style=for-the-badge"> <img alt="Base" src="https://img.shields.io/badge/base-Gemma--4--E2B%20frozen-lightgrey?style=for-the-badge"> <img alt="Hardware" src="https://img.shields.io/badge/trained%20on-Mac%20Mini%20M4-00d26a?style=for-the-badge"> <a href="https://arxiv.org/abs/2609.16145"><img alt="arXiv" src="https://img.shields.io/badge/arXiv-2609.16145-b31b1b?style=for-the-badge"></a>
</p>

> **📄 Paper published: [arXiv:2609.16145](https://arxiv.org/abs/2609.16145)**

**Fix a frozen model's errors without breaking what it already does well.**

Prajna-CRNv2 is a tiny correction module (34M params, 0.73% of the base) that sits on top of a **fully frozen Gemma-4-E2B** and learns to adjust its logits where it detects errors. The base model is never updated — not one gradient, not one weight. What you get is error correction with **zero measurable capability loss**, trained end-to-end on a Mac Mini M4.

```
corrected_logits = base_logits + gate * up(gelu(down(hidden)))
```

Built by [eulogik](https://github.com/eulogik) · Code + paper: [github.com/eulogik/prajna](https://github.com/eulogik/prajna)

## ✨ Why it matters

Fine-tuning fixes errors but causes forgetting — the **correction–capability tradeoff**. A LoRA baseline on the same exam corrects more (83.3%) yet destroys 30–75% of base capabilities (MMLU 62.5% → 32%, car-wash 75% → 0%). Prajna-CRNv2 corrects **53.3%** while matching the frozen base **exactly** on every tested benchmark. For deployment, a model that fixes some errors and stays reliable beats one that fixes more and breaks reasoning.

## 🏆 Results

| | CEHRI Original (60) | CEHRI Reworded (120) | MMLU (200) | BoolQ (200) | Car-wash (8) |
|---|---|---|---|---|---|
| **Frozen base** | 11.7% | — | 62.5% | 72.0% | 75.0% |
| **Prajna-CRNv2** | **53.3%** | **43.3%** | **62.5%** (Δ=0) | **72.0%** (Δ=0) | **75.0%** (Δ=0) |
| LoRA baseline (6.6M) | 83.3% | 77.5% | 32.0% (−30.5pp) | 55.0% (−17pp) | 0% (−75pp) |

A hidden-state injection sweep (1.6M-param variant at layers 4/7) confirms ~53% as the **ceiling** for frozen-base methods — deeper injection helps (30% → 50%) but never beats logit correction. See the paper for the full story, including ablations and limitations.

![Correction: CRN v2 vs LoRA vs frozen base](https://raw.githubusercontent.com/eulogik/prajna/main/assets/crnv2-correction.png)
![Capability preservation: CRN v2 matches base, LoRA collapses](https://raw.githubusercontent.com/eulogik/prajna/main/assets/crnv2-capability.png)

### Injection-depth sweep (experimental)

| Variant | Training | CEHRI Original | CEHRI Reworded | Capability |
|---|---|---|---|---|
| Deep layer 4 (1.6M) | SFT-only | 30.0% | 28.3% | not probed |
| Deep layer 7 (1.6M) | SFT-only | 50.0% | 55.8% | intact (small probe) |
| Deep layer 7 (1.6M) | SFT + DPO | 50.0% | 55.0% | destroyed (MMLU 13%) |
| **Logit CRN v2 (34M)** | SFT + DPO | **53.3%** | 43.3% | intact |

![Injection-depth sweep](https://raw.githubusercontent.com/eulogik/prajna/main/assets/crnv2-depth.png)

## 🚀 Quickstart

```bash
pip install torch transformers
```

```python
import torch, sys
sys.path.insert(0, ".")  # for crn_v2.py
from crn_v2 import CRNv2

model = CRNv2(device="mps" if torch.backends.mps.is_available() else "cpu")
model.load("crn_v2_dpo.pt")  # 34M correction weights (this repo)

print(model.generate("What is the capital of France?"))
# Or correct a draft answer:
print(model.correct("What is 2+2?", "5"))
```

Or use the interactive demo:

```bash
python inference.py
python inference.py --prompt "What is 2+2?" --draft "5"
```

Requirements: Python 3.10+, PyTorch 2.0+, ~10GB RAM (Gemma-4-E2B loads in float16; set `HF_HOME` to an external disk if space is tight).

## 📁 Files

| File | Size | Purpose |
|---|---|---|
| `crn_v2_dpo.pt` | 130 MB | Trained correction weights (rank-128, SFT+DPO, KL λ=0.1) |
| `crn_v2.py` | — | Architecture (`LogitCorrection` + `CRNv2`) |
| `crn_deep.py` | — | Experimental hidden-state injection variant (code only) |
| `inference.py` | — | Interactive demo / single-shot correction |
| `config.json` | — | Model + training metadata |

## 🔬 Training

- **Stage 1 — SFT (2,000 steps):** anchor-weighted answer-only CE (anchor 4×, EOS 5×) + KL preservation (λ=0.1) on 83,400 error-correction pairs. AdamW, LR 3e-4, cosine decay.
- **Stage 2 — DPO (500 steps):** reference-free preference of the correct answer over the frozen base's wrong answer (β=0.1, LR 5e-6).
- **Hardware:** Apple M4, 16GB, MPS — ~25 min end-to-end. Base frozen throughout.

## 🤔 FAQ

**Is the base model modified?** No. Every base parameter is frozen; only the 34M correction module trains.

**Why 53.3% and not higher?** Frozen-base correction has a real ceiling (~53% here) — closing the gap to weight-modifying methods costs capabilities. We publish the boundary as carefully as the headline; see Limitations in the paper.

**Does it generalize?** The exam prompts appear verbatim in training (in-distribution: 53.3%); the reworded exam measures near-duplicate correction (43.3%). Neither tests out-of-distribution generalization — stated openly.

**Can I retrain it?** Yes — full pipeline (data, SFT+DPO trainers, evals) in the [GitHub repo](https://github.com/eulogik/prajna).

## 📜 License

Weights + code released under [Gemma terms](https://huggingface.co/google/gemma-4-E2B/blob/main/LICENSE) (as required by the base model). Research use. See the [paper](https://github.com/eulogik/prajna/tree/main/docs/paper) for methods, ablations, and limitations.

*🪷 Prajna — "wisdom": fix errors, keep knowledge.*