---
base_model: Qwen/Qwen2.5-Coder-3B-Instruct
library_name: peft
tags:
  - lora
  - peft
  - hallucination-detection
  - venra
license: apache-2.0
---

# VeNRA — LoRA Adapter 

Fine-tuned LoRA adapter on `Qwen/Qwen2.5-Coder-3B-Instruct` for
hallucination detection in RAG pipelines.

## Available Adapters

| Branch | Rank | Description |
|--------|------|-------------|
| `r96`  | 96   | Lighter, faster inference |
| `r128` | 128  | Higher capacity |

## Labels
- `Found` — supported by context
- `General`   — common knowledge
- `Fake` — contradicts or unsupported by context

## Usage
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

BASE_MODEL = "Qwen/Qwen2.5-Coder-3B-Instruct"

# Load r96
model_r96 = PeftModel.from_pretrained(base, "pagand/venra", revision="r96")

# Load r128
model_r128 = PeftModel.from_pretrained(base, "pagand/venra", revision="r128")

# Pinned to a specific snapshot tag
model = PeftModel.from_pretrained(model, "pagand/venra", revision="r96-v1.0")
```

## Training Details
- Rank: 96/128
- Learning rate: 1e-4
- Weight decay: 0.10
- Training regime: WeightedLabelTrainer