---
library_name: transformers
license: apache-2.0
language:
  - en
base_model: google/gemma-3-1b-it
tags:
  - gemma
  - finetune
  - qlora
  - chatbot
  - tars
---

# Model Card for TARS (Gemma 3 1B Fine-tune)

This is a fine-tuned version of `google/gemma-3-1b-it` trained to act as the **TARS astronaut assistant** from *Interstellar*.  
It is designed to be professional for tasks but witty for off-topic chat, and its responses are guided by a simulated user emotion tag.

---

## Model Details

### Model Description

This model is a QLoRA fine-tune of `google/gemma-3-1b-it` on a custom synthetic dataset.  
The goal was to create a chatbot that embodies the **TARS persona**:

- **Task-Oriented:** Professional, direct, and helpful for mission-related queries.  
- **Persona-Driven:** Witty, empathetic, or humorous for off-topic or personal chat.  
- **Emotion-Aware:** The model's response style is influenced by a `[Detected Emotion: ...]` tag.  

**Developed by:** (huggingface.co/am-om)  
**Shared by:** (Om Singh)  
**Model type:** Causal Language Model  
**Language(s):** English (`en`)  
**License:** apache-2.0  
**Finetuned from model:** `google/gemma-3-1b-it`

---

## Model Sources (optional)

- **Repository:** [https://huggingface.co/am-om/tars_ai]

---

## Uses

### Direct Use

This model is intended for **direct use as a chatbot**, following a specific prompt format.

⚠️ **Important:** This model requires a specific prompt format that includes a detected emotion.  
Do **not** send raw text as the user query.



#### Prompt Format

The user turn *must* follow this structure:

```
[Detected Emotion: {emotion}]
[User Query: {your_text_here}]
```

**Example:**

```
[Detected Emotion: anxious]
[User Query: Are we going to make it?]
```

### Out-of-Scope Use

This model is not intended for:
* Any use without the required `[Detected Emotion: ...]` and `[User Query: ...]` tags.
* Use as a base model for further fine-tuning.
* Any critical decision-making without human oversight.

## How to Get Started with the Model

Use the code below to get started with the model.

```python
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import torch

# Load the model from the Hub
model_id = "am-om/tars_ai" 

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

pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer
)

# --- Define your chat history ---
# The system prompt is automatically loaded from the tokenizer's chat template.
messages = []

# Example query
user_query = "I'm feeling a bit lonely out here."
emotion = "sad"

# Format the input correctly!
formatted_input = f"[Detected Emotion: {emotion}]\n[User Query: {user_query}]"

messages.append({"role": "user", "content": formatted_input})

# --- Generate the response ---
prompt = pipe.tokenizer.apply_chat_template(
    messages, 
    tokenize=False, 
    add_generation_prompt=True
)

outputs = pipe(
    prompt,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
    pad_token_id=pipe.tokenizer.eos_token_id
)

# Extract and print just the new response
response = outputs[0]["generated_text"][len(prompt):].strip()
print(f"TARS: {response}")
```

## Training Details

### Training Data

This model was fine-tuned on a custom, synthetically-generated dataset of 344 prompt/response pairs. The dataset was designed to teach the model to differentiate between task-oriented and persona-driven queries based on the emotion tag.

### Training Procedure

The model was fine-tuned using QLoRA for 3 epochs. The adapter (from checkpoint-156, the best-performing epoch) was then merged with the base model.

#### Training Hyperparameters

* **Framework:** TRL (Transformer Reinforcement Learning)
* **Quantization:** 4-bit (bnb_4bit_quant_type="nf4")
* **LoRA `r`:** 16
* **LoRA `alpha`:** 32
* **LoRA `dropout`:** 0.05
* **Optimizer:** paged_adamw_8bit
* **Learning Rate:** 5e-5
* **LR Scheduler:** constant
* **Epochs:** 3
* **Batch Size:** 4

## Environmental Impact

Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).

* **Hardware Type:** NVIDIA T4  
* **Hours used:** ~4hours
* **Cloud Provider:** Google Colab  
* **Compute Region:** (e.g., us-central1 - *check your Colab instance*)  
* **Carbon Emitted:** ~5.5 g CO2eq (Estimated)

## Technical Specifications [optional]

### Model Architecture and Objective

This is a standard decoder-only Transformer (Gemma 3) fine-tuned with a Causal Language Modeling objective.

### Compute Infrastructure

#### Hardware
* NVIDIA T4 16GB (Google Collab )

#### Software
* `transformers`
* `trl`
* `bitsandbytes`
* `accelerate`
* `peft`

## Model Card Authors [optional]

(Om Singh)(huggingface.co/am-om)

## Model Card Contact

(huggingface.co/am-om)