---
base_model: unsloth/Qwen2.5-3B-Instruct-bnb-4bit
library_name: transformers
tags:
- unsloth
- trl
- qlora
- astrology
- vedic-astrology
- hindi
- hinglish
- lora
license: apache-2.0
---

# Vedaz: AI Vedic Astrologer

Vedaz is a fine-tuned LoRA adapter on top of `unsloth/Qwen2.5-3B-Instruct-bnb-4bit`, designed to provide **compassionate**, **empathetic**, and **balanced** Vedic astrological guidance.  
It focuses on sensitive life areas such as career, relationships, finance, health, and spiritual growth, while maintaining clear ethical boundaries and avoiding fatalistic or harmful predictions.

---

## Model Details

- **Developed by:** FinetunerHegde  
- **Model type:** Causal Language Model (LoRA fine-tuned adapter)  
- **Base model:** Qwen2.5-3B-Instruct (4-bit, via Unsloth)  
- **Languages:** English, Hindi, Hinglish (code-mixed)  
- **Training framework:** Unsloth + TRL (QLoRA)  
- **Intended domain:** Vedic astrology chat assistant  
- **Input style:** Instruction/chat format (system + user messages)

### Model Motivation

The goal of Vedaz is to make Vedic astrology guidance more accessible, gentle, and practical for everyday life decisions.  
Instead of deterministic predictions, the model encourages reflection, self-awareness, and constructive actions aligned with the user’s context.

---

## Quick Start (Unsloth Inference)

Vedaz is a LoRA adapter and should be loaded with the base model using the `unsloth` library for best performance and low memory usage.

```python
from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "FinetunerHegde/vedaz",
    max_seq_length = 2048,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

messages = [
    {
        "role": "system",
        "content": (
            "You are Vedaz, an AI Vedic astrologer. "
            "You give compassionate, balanced guidance based on Vedic principles. "
            "You avoid absolute predictions and encourage practical, ethical actions."
        ),
    },
    {
        "role": "user",
        "content": "Meri shaadi kab hogi? DOB: 15 August 1996, 7:00 AM, Pune.",
    },
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to("cuda")

outputs = model.generate(
    input_ids=inputs,
    max_new_tokens=300,
    temperature=0.7,
    do_sample=True,
)

print(tokenizer.decode(outputs[inputs.shape[-1]:], skip_special_tokens=True))
```

---

## Recommended Prompting Style

For best results:

- Provide **DOB, time, and place** of birth in a single message.  
- Mention the **area of concern** (career, marriage, finance, health, spiritual growth, etc.).  
- Use natural language in English, Hindi, or Hinglish.

Example prompts:

- `"DOB: 02 Jan 1998, 3:45 PM, Bengaluru. Career-wise next 2–3 years ka overview batao."`  
- `"Meri relationship ke bare mein guidance chahiye. DOB: 10 March 1995, 9:15 AM, Delhi."`

---

## Intended Uses

### Direct Use

- Vedic astrology–style chat assistant for:
  - High-level life guidance and reflection  
  - Relationship and career counseling in a gentle, non-fatalistic manner  
  - Spiritual and emotional support aligned with Vedic concepts  

### Downstream Use

- Integrating Vedaz into:
  - Telegram/WhatsApp/Discord bots  
  - Web apps (e.g., Gradio/Streamlit frontends)  
  - Personal assistants or dashboards that offer astrology-themed insights

### Out-of-Scope Use

Vedaz **must not** be used for:

- Medical diagnosis, emergency advice, or mental health treatment  
- Financial, legal, or investment decisions where professional advice is required  
- Absolute predictions about death, accidents, or harmful outcomes  
- Any form of discrimination, hate speech, or harassment

---

## Bias, Risks, and Limitations

- The model is trained on curated chat data with Vedic astrology–style content and may reflect biases present in the training set.  
- It can generate culturally specific interpretations which may not align with all belief systems or personal philosophies.  
- Outputs are **not** guaranteed to be accurate, and should be treated as guidance or reflection, not as “truth”.

### Recommendations

- Always treat the model as a **supportive conversational tool**, not an authority.  
- Cross-check important decisions with qualified professionals (career counselors, doctors, financial advisors, etc.).  
- Avoid using the model with vulnerable users without proper supervision or safety layers.

---

## Training Details

> Note: The following is a high-level overview. Update with exact numbers if available.

### Training Data

- Custom chat dataset focused on:
  - Vedic astrology–style Q&A  
  - Conversational guidance on life topics  
  - Hindi, English, and Hinglish mixed dialogues  
- Sensitive topics (health, relationships, finance) labeled or filtered to encourage empathetic and safe responses.

### Training Procedure

- **Method:** QLoRA fine-tuning on `unsloth/Qwen2.5-3B-Instruct-bnb-4bit`  
- **Frameworks:** Unsloth + TRL  
- **Objective:** Instruction-following and chat-style responses with strong safety and empathy constraints.

### Example Hyperparameters (to customize)

- LoRA rank and alpha tuned for compactness  
- Max sequence length: 2048 tokens  
- Optimization target: fast inference on consumer GPUs using 4-bit quantization.

---

## Evaluation

> This section is indicative. Fill in with your actual evaluation setup and metrics when available.

- **Qualitative evaluation:**  
  - Manually reviewed responses for empathy, safety, and adherence to Vedic framing.  
  - Checked behavior on sensitive queries (marriage, job loss, health anxiety).

- **Key focus areas:**  
  - Non-fatalistic guidance  
  - Constructive, actionable suggestions  
  - Avoidance of explicit harmful content

---

## Technical Specifications

### Architecture

- Base: Qwen2.5-3B-Instruct (decoder-only transformer)  
- Adapter: LoRA layers via Unsloth (QLoRA on 4-bit base model)  
- Max context: 2048 tokens (recommended)

### Hardware & Inference

- Optimized for:
  - Single consumer GPU (e.g., 8–12 GB VRAM)  
  - Fast inference with 4-bit quantization  
- Compatible with:
  - `unsloth` API for loading and generation  
  - Standard Hugging Face ecosystem via adapters

---

## Ethical Considerations

- Vedaz is designed to **support** users emotionally, not to control or dictate life choices.  
- It should always:
  - Encourage self-responsibility and practical steps.  
  - Avoid deterministic or fear-based predictions.  
  - Respect user autonomy and diversity of beliefs.

If you deploy Vedaz publicly, please clearly mention that it is an AI system and not a certified astrologer, therapist, or advisor.

---

## Citation

If you use Vedaz in academic work, projects, or demos, you can cite it as:

```bibtex
@misc{vedaz2026,
  title        = {Vedaz: AI Vedic Astrologer (LoRA on