🌌 Gemma-2b-TARS-SFT: Technical Model Card

Fine-tuned Gemma-2-2B-it optimized for Creative Writing, Technical Assistance, and Distinctive Persona-Driven Chat.

Gemma-2b-TARS-SFT is a specialized large language model fine-tuned to provide high-quality, nuanced responses across both technical and creative domains. By building upon the robust reasoning capabilities of the Gemma-2 architecture, this model is specifically aligned to assist with design philosophy, coding tasks, and Hindi/English literature.

---

🎭 Model Persona & Roleplay

Unlike the standard, sterile base model, TARS has been fine-tuned with a distinct, slightly sarcastic, and theatrical personality (heavily inspired by science-fiction tropes).

• Emotes: The model may spontaneously use action tags (e.g., *Adjusts welding goggles* or *Leans in conspiratorially*).
• Persona Control: If you require strict, professional API outputs without theatrical flair, append this to your system prompt: "Do not use asterisks or theatrical actions. Provide only the direct, professional answer."

---

🛠 Model Details

• Base Model: google/gemma-2-2b-it
• Architecture: 2.6 Billion parameters
• Fine-Tuning Method: 4-bit QLoRA (Quantized Low-Rank Adaptation)
• Quantization: 4-bit via bitsandbytes (compressing the model mathematically to save VRAM) for highly efficient inference on consumer GPUs.
• Creator: prash616 (Prashant)

---

📊 Training Procedure & Data

The model was developed using a Supervised Fine-Tuning (SFT) strategy. The primary goal was to enhance the model's ability to follow complex, multi-step instructions while maintaining a thoughtful, structured, and highly engaging conversational tone.

1. Datasets

• Databricks Dolly-15k: Utilized to build a strong foundation in general instruction-following, brainstorming, classification, and open QA tasks.
• Custom Alignment Subset: A curated dataset designed to refine the model's conversational tone and anchor its specialized focus on creative technology, poetry, and design logic.

2. Training Hyperparameters

Parameter	Value
Learning Rate	1e-4
LoRA Rank (r)	16
LoRA Alpha	32
Max Steps	300
Optimiser	AdamW (8-bit)
Target Modules	q, k, v, o, gate, up, down

---

🚀 Usage & Implementation (Google Colab / Python)

Prerequisites: Because this model is based on Gemma-2, you must have a Hugging Face token and accept the official Google Gemma terms of service.

1. Install Optimized Libraries:

pip install --no-deps unsloth unsloth_zoo "xformers<0.0.29" "trl<0.9.0" peft accelerate bitsandbytes
import torch
import getpass
from unsloth import FastLanguageModel

# Secure Token Input
hf_token = getpass.getpass("Enter your Hugging Face Token: ")

# Load the model in highly-efficient 4-bit mode (Critical for free-tier GPUs)
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "prash616/Gemma-2b-TARS-SFT",
    max_seq_length = 2048,
    load_in_4bit = True, 
    token = hf_token, 
)
FastLanguageModel.for_inference(model) # Enables 2x faster generation

# Format the Prompt
messages = [
    {"role": "system", "content": "You are TARS, an AI assistant specialized in creative technology and literature. You were created by Prashant."},
    {"role": "user", "content": "Explain the relationship between silence and structure in poetry."}
]

# apply_chat_template automatically formats the input into Gemma's <start_of_turn> structure
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True, 
    add_generation_prompt=True, 
    return_tensors="pt",
    return_dict=True # Generates both input_ids and attention_mask
).to("cuda")

# Generate the Response
outputs = model.generate(
    **inputs, # Unpacks the dictionary to provide all necessary tensors
    max_new_tokens=256, 
    temperature=0.7,
    do_sample=True
)

print("\n--- TARS RESPONDS ---\n")
print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])