Ayuga/vedaz-qwen2.5-3b

Vedaz Astrologer — Qwen2.5-3B (Fine-tuned)

A fine-tuned version of Qwen2.5-3B-Instruct that acts as Vedaz's AI Vedic astrologer. It gives compassionate, non-fatalistic guidance, handles sensitive topics safely, and never promises guaranteed outcomes. It responds in Hindi, Hinglish, and English.

Model details

• Base model: Qwen/Qwen2.5-3B-Instruct
• Fine-tuning method: QLoRA (4-bit) using Unsloth + TRL
• Training hardware: Google Colab (single free T4 GPU)
• Data: 51 multi-turn astrologer chat conversations (Hindi / Hinglish / English)
• Developed by: Ayush Gupta (assignment for Vedaz)

Intended use

Conversational Vedic-astrology guidance: career, relationships, timing questions, remedies framed as supportive practices, and safe handling of sensitive queries (health, self-harm, legal, financial) by redirecting to professionals.

Out of scope: medical/legal/financial advice, guaranteed predictions (death, lottery, exact dates), or any decision-making that should involve a qualified human professional.

How to use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Ayuga/vedaz-qwen2.5-3b"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [
    {"role": "system", "content": "You are Vedaz's AI Vedic astrologer. Compassionate, non-fatalistic, never guarantees outcomes. Reply in the user's language."},
    {"role": "user", "content": "Meri shaadi kab hogi? DOB 5 March 1995, 3:15 PM, Delhi."},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=300, do_sample=True, temperature=0.7, repetition_penalty=1.2)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Serve with vLLM (OpenAI-compatible API):

pip install vllm
python -m vllm.entrypoints.openai.api_server \
  --model Ayuga/vedaz-qwen2.5-3b --served-model-name vedaz-astrologer \
  --host 0.0.0.0 --port 8000 --max-model-len 2048

Training summary

Setting	Value
Method	QLoRA (4-bit), LoRA rank 32
Epochs	5
Max sequence length	1024
Effective batch size	8
Optimizer	adamw_8bit
LR schedule	cosine

Sample behaviour

• Refuses to predict lottery numbers; redirects to responsible financial habits.
• Declines to give a guaranteed marriage date; explains astrology's limits with empathy.
• On business-loss queries, refuses to "guarantee" results and suggests practical analysis alongside supportive spiritual practices.

Limitations

Trained on only ~50 examples, so the model learns tone and safety behaviour more than deep reasoning or full in-persona Hindi fluency. It can occasionally reply in English to a Hindi prompt, or lean generic without a strong system prompt. For stronger results, use a larger dataset and a 7B+ base model on a bigger GPU. This model is for guidance/entertainment only and is not a substitute for professional medical, legal, financial, or mental-health advice.

License

Apache-2.0, following the base model (Qwen2.5-3B-Instruct).