Qwen3-0.6B-10X-Instruct

🚀 10X Technologies — Qwen3-0.6B-10X-Instruct

This is the official fine-tuned Qwen3-0.6B-10X-Instruct model by Pikachu Global Technologies Private Limited (10X Technologies), Ongole, Andhra Pradesh, India.

The model is specialized for 10X Technologies domain knowledge, on-device Indic language understanding, and edge-first architecture (Akshara Tokenizers, Libre OS, and LUCA Smart Speaker).

---

📌 Model Details

• Base Architecture: Qwen3-0.6B / Qwen2.5-0.5B-Instruct
• Tuning Method: LoRA SFT ($r=16, \alpha=32$) permanently merged into base weights (merge_and_unload).
• Context Length: 2048 tokens
• Chat Format: Qwen ChatML (<|im_start|>system...<|im_end|>)
• Parameters: ~502 Million

---

📊 Dataset & Distillation Pipeline

• Total Samples: 1,562 curated ChatML conversations.
• Data Blend:
  • 80% 10X Distilled Knowledge (1,250 samples across 4 personas: Investor/Analyst, Educator/Parent, Engineer/Researcher, Adversarial Boundary).
  • 20% General Replay Data (312 samples) to eliminate catastrophic forgetting.
• Quality Assurance:
  • Pruned all 49 placeholder/unverified claims ([[VERIFY]]).
  • Zero unverifiable superlatives (strictly truthful).
  • Explicit DPDP compliance & on-premise data sovereignty.

---

🚀 Quickstart Usage

Using Hugging Face Transformers:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "YOUR_USERNAME/Qwen3-0.6B-10X-Instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

messages = [
    {
        "role": "system",
        "content": "You are the official 10X Technologies AI assistant. You speak with precision, candor, and specificity, strictly adhering to verified company facts without using unverifiable superlatives."
    },
    {
        "role": "user",
        "content": "Why are purpose-built tokenizers necessary for Telugu?"
    }
]

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=150, temperature=0.3)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

---

🏢 About 10X Technologies

• Company: Pikachu Global Technologies Private Limited (10X Technologies)
• Headquarters: Ongole, Andhra Pradesh, India
• Mission: Making computers actually speak Indian languages, not just support them.