sensix-zo/sensix-paite-4b-instruction-16bit

Sensix Paite 4B Instruction (16-bit)

This model is a fine-tuned version of Gemma 3 4B developed through a two-stage training pipeline: Continued Pre-Training (CPT) for vocabulary acquisition and Supervised Fine-Tuning (SFT) for instruction following. It is optimized for native Paite linguistic reasoning.

Model Details

• Base Model: unsloth/gemma-3-4b-it
• Target Language: Paite (pck)
• Precision: 16-bit bfloat16 (Full Precision)
• Training Method: LoRA with Hard Merge
• Framework: Unsloth, TRL, and PEFT

Training Procedure

Stage 1: Continued Pre-Training (CPT)

Knowledge injection was performed using the PERFECT_PAITE_DATA.jsonl dataset (articles, news, and long-form paragraphs).

• Learning Rate: 2e-4
• LoRA Config: r=64, alpha=128
• Focus: Modern Paite vocabulary expansion, excluding repetitive scriptural fragments.

Stage 2: Supervised Fine-Tuning (SFT)

The model was refined on mixed_alpaca_paite_2026-04-09.jsonl to establish instruction-following logic.

• Learning Rate: 2e-5
• Epochs: 3
• Data Packing: Enabled
• Prompt Format: Gemma 3 Chat Template (Messaging Format)

Technical Implementation: Hard Merge Strategy

To prevent the common weight-scrambling issue (known as the "Calcium/Blades" gibberish bug) found in Gemma 3/4 merges, this model was fused using an official Hard Merge (PEFT merge_and_unload) rather than simple weight averaging. This ensures 100% stability and preserves the model's reasoning logic.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "sensix-zo/sensix-paite-4b-instruction-16bit"

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

# Example Prompt
messages = [
    {"role": "user", "content": "Paite pau hi bangchiah in a poimoh hiam?"}
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Limitations

This model is optimized for the Paite language. While it retains English capabilities, users should verify complex technical outputs for accuracy.