README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-1.7B
tags:
  - pathi-ai
  - pathi-labs
  - pathi-lite-instruct-1.7B
  - qwen3
  - qwen
  - causal-language-model
  - text-generation
  - conversational
  - instruction-tuning
  - lora
  - peft
  - fine-tuned
language:
  - en
library_name: peft
pipeline_tag: text-generation
---

<div align="center">

# Pathi-Lite-Instruct-1.7B

**A lightweight, instruction-tuned language model by [Pathi Labs](https://www.pathilabs.com)**

[![Website](https://img.shields.io/badge/Website-pathilabs.com-7A5AF8)](https://www.pathilabs.com)
[![License](https://img.shields.io/badge/License-Apache%202.0-blue)](https://www.apache.org/licenses/LICENSE-2.0)
[![Base Model](https://img.shields.io/badge/Base-Qwen3--1.7B-orange)](https://huggingface.co/Qwen/Qwen3-1.7B)

</div>

---

## Model Overview

**Pathi-Lite-Instruct-1.7B** is an instruction-tuned language model developed by **Pathi Labs LLP**, fine-tuned from [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) using **Low-Rank Adaptation (LoRA)**. It is designed to deliver efficient, high-quality instruction-following performance while remaining lightweight enough for accessible deployment.

This release is part of Pathi Labs' ongoing work in applied AI research and lightweight model development.

| | |
|---|---|
| **Developed by** | [Pathi Labs LLP](https://www.pathilabs.com) |
| **Model type** | Causal decoder-only transformer (instruction-tuned) |
| **Base model** | [Qwen/Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) |
| **Fine-tuning method** | LoRA (Low-Rank Adaptation) |
| **Language(s)** | English |
| **License** | Apache 2.0 |
| **Contact** | [Info@pathilabs.com](mailto:Info@pathilabs.com) |

---
### Intended Use

**Primary use cases:**
- Instruction following and general-purpose conversational assistance
- Lightweight deployment in resource-constrained environments
- A base for further fine-tuning or research experimentation

**Out-of-scope use:**
- High-stakes decision-making (medical, legal, financial) without human oversight
- Generation of harmful, misleading, or illegal content
- Use cases requiring guarantees of factual accuracy

### Limitations

- As a 1.7B-parameter model, it has a smaller knowledge and reasoning capacity than larger frontier models.
- May produce inaccurate, incomplete, or biased outputs; outputs should be reviewed before use in production.
- LoRA fine-tuning adapts behavior but does not remove limitations inherited from the base model.

---

## How to Use

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "pathilabs/Pathi-Lite-Instruct-1.7B"

# 1. Load Tokenizer and Merged Model
tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=True
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16, # bfloat16 matches native Qwen precision perfectly
    device_map="auto",
    trust_remote_code=True
)

# 2. Format inputs using the required Chat Template
messages = [
    {"role": "user", "content": "What is artificial intelligence?"}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(
    text,
    return_tensors="pt"
).to(model.device)

# 3. Generate response with clean configuration parameters
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=True,
        temperature=0.7,
        top_p=0.8
    )

# 4. Decode output cleanly
response = tokenizer.decode(
    outputs[0][len(inputs.input_ids[0]):], # Cuts out the input prompt from printing twice
    skip_special_tokens=True
)

print(response)
```

---
### Model Training Method

Fine-tuning was performed using **LoRA (Low-Rank Adaptation)**, a parameter-efficient technique that freezes the base model weights and trains small injected rank-decomposition matrices in select layers.

# Model Details

| Property | Value |
|---|---|
| Model name | Pathi-Lite-Instruct-1.7B |
| Organization | Pathi Labs LLP |
| Base model | Qwen/Qwen3-1.7B |
| Model family | Qwen3 |
| Model type | Causal Language Model |
| Parameter scale | ~1.7B |
| Fine-tuning method | LoRA / PEFT |
| Final release format | Merged model |
| Framework | Hugging Face Transformers |
| Serialization | Safetensors |
| Primary task | Text generation |
| Intended language | English |
| Developer | Pathi Labs LLP |

---
## Responsible Use

Pathi Labs encourages responsible deployment of this model. Users should:
- Evaluate outputs for accuracy and safety before use in production systems
- Avoid deploying the model in high-stakes domains without human review
- Respect the licensing terms of both this model and the base Qwen3-1.7B model

---

## Citation

If you use this model in your work, please cite:

```bibtex
@misc{pathilite2026,
  title  = {Pathi-Lite-Instruct-1.7B},
  author = {Pathi Labs LLP},
  year   = {2026},
  url    = {https://huggingface.co/pathilabs/Pathi-Lite-Instruct-1.7B},
  note   = {Fine-tuned from Qwen/Qwen3-1.7B using LoRA}
}
```

---

## About Pathi Labs

**Pathi Labs LLP** is an AI research and development company building applied machine learning solutions.

- 🌐 Website: [www.pathilabs.com](https://www.pathilabs.com)
- 📧 Contact: [Info@pathilabs.com](mailto:Info@pathilabs.com)

---

## Acknowledgements

This model is built on top of [Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) by the Qwen Team, Alibaba Cloud. We thank the Qwen team for releasing their models openly.