README.md

---
base_model: microsoft/Phi-4-mini-instruct
library_name: peft
pipeline_tag: text-generation
tags:
  - lora
  - peft
  - phi-4
---

# davanai 2

A LoRA adapter fine-tuned on **microsoft/Phi-4-mini-instruct** (3.8B), developed by **Vega Aiden Lab**.

## Model details

- **Developed by:** Vega Aiden Lab
- **Model:** davanai 2
- **Base model:** microsoft/Phi-4-mini-instruct
- **Method:** LoRA (PEFT), rank 16

---

## Run it locally — step by step

> This is a LoRA adapter, not a full model. You download the base model **microsoft/Phi-4-mini-instruct** and apply this adapter on top. Phi-4-mini is small (3.8B), so it runs on a normal GPU (8 GB+) or even on CPU if you're patient.

### 1. Install the libraries

```bash
pip install torch transformers peft accelerate safetensors
```

### 2. Log in to Hugging Face

```bash
pip install -U "huggingface_hub[cli]"
hf auth login
```

Paste a token from https://huggingface.co/settings/tokens when asked.

### 3. Create a file called `run.py`

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "microsoft/Phi-4-mini-instruct"
adapter = "emmaoba/davanai-2"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base,
    device_map="auto",
    torch_dtype="auto",
)

# Apply the davanai 2 adapter
model = PeftModel.from_pretrained(model, adapter)
model.eval()

messages = [{"role": "user", "content": "Hello! Who are you?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, do_sample=True, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```

### 4. Run it

```bash
python run.py
```

The first run downloads the base model (once, then cached), then prints davanai 2's reply.

---

## Training configuration

- Method: LoRA (PEFT)
- Rank (r): 16
- Alpha: 32
- Dropout: 0.05
- Target modules: qkv_proj, o_proj, gate_up_proj, down_proj