README.md

---
base_model: microsoft/phi-3-mini-128k-instruct
library_name: peft
model_name: phi3-nl2bash-lora
tags:
- lora
- nl2bash
- sft
- bash
- transformers
- trl
license: mit
pipeline_tag: text-generation
---

# phi3-nl2bash-lora

This repository contains **LoRA adapter weights** fine-tuned on the
[`jiacheng-ye/nl2bash`](https://huggingface.co/datasets/jiacheng-ye/nl2bash)
dataset to convert **natural language instructions into Linux bash commands**.

> ⚠️ This repository contains **LoRA adapters only**, not the base model.  
> You must load these adapters on top of  
> **`microsoft/phi-3-mini-128k-instruct`**.

---

## Intended use

The model is trained to output **only valid bash commands**, with no explanations.

**Example**

Input:
```
List all .txt files recursively and count lines
```

Output:
```bash
find . -name "*.txt" | xargs wc
```

## Training summary
- Base model: microsoft/phi-3-mini-128k-instruct
- Fine-tuning method: LoRA (PEFT)
- Trainer: TRL SFTTrainer
- Dataset: jiacheng-ye/nl2bash
- Output format: Bash commands only

## Loading example
```
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model = "microsoft/phi-3-mini-128k-instruct"
lora_model = "ayertiam/phi3-nl2bash-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

model = PeftModel.from_pretrained(model, lora_model)
model.eval()
```

## Notes
- These adapters are model-specific and only compatible with `microsoft/phi-3-mini-128k-instruct`.
- For Ollama or GGUF usage, the LoRA must be merged into the base model and converted before inference as done here https://huggingface.co/ayertiam/phi3-nl2bash-gguf