lora_backbone/README.md

---
base_model: google/gemma-4-E4B-it
library_name: peft
tags:
  - base_model:adapter:google/gemma-4-E4B-it
  - lora
  - peft
  - classone
  - decision-model
---

# LoRA Adapter — classone-gemma4-e4b

PEFT LoRA adapter for **[devops-thiago/classone-gemma4-e4b](https://huggingface.co/devops-thiago/classone-gemma4-e4b)**, the ClassOne System 1 decision model built on [`google/gemma-4-E4B-it`](https://huggingface.co/google/gemma-4-E4B-it).

This subfolder holds only the low-rank adapter weights. The merged standalone weights live at the repository root, and the Noul / Choice / Score decision heads live in `classone_heads.pt` — load the heads alongside whichever backbone form you use.

## Adapter Configuration

| Setting | Value |
|---|---|
| Base model | `google/gemma-4-E4B-it` |
| PEFT type | LORA |
| Rank (r) | 16 |
| `lora_alpha` | 32 |
| `lora_dropout` | 0.05 |
| Bias | none |
| Target modules | `.*language_model.*(q_proj\|o_proj\|gate_proj\|up_proj\|down_proj).*` |
| PEFT version | 0.21.0 |

## Usage

```python
from peft import PeftModel
from transformers import AutoModel

base = AutoModel.from_pretrained("google/gemma-4-E4B-it")
model = PeftModel.from_pretrained(base, "devops-thiago/classone-gemma4-e4b", subfolder="lora_backbone")
```

For end-to-end decision inference (packing questions, calibrated outputs), see the [model card](https://huggingface.co/devops-thiago/classone-gemma4-e4b).

## Attribution & Legal

- Base model: [google/gemma-4-E4B-it](https://huggingface.co/google/gemma-4-E4B-it) (Google) — Apache License 2.0
- Architecture: [devops-thiago/class-one](https://github.com/devops-thiago/class-one) — Apache 2.0