README.md

## library\_name: transformers tags: [scientific, instruction-following, gemma, lora, gemma3-270m, science]

# Model Card for ps2program/gemma3-270m-scisinstruct

This is a **LoRA-fine-tuned version of Gemma 3 (270M)**, specialized for **scientific instruction-following and reasoning tasks**. The model has been trained on the `zd21/SciInstruct` dataset to excel at generating explanations, summaries, and scientific reasoning outputs in response to natural language prompts.

## Model Details

### Model Description

This model extends the Google-developed Gemma 3 base language model using LoRA (Low-Rank Adaptation) adapters. This technique allows for efficient fine-tuning on a specific task without modifying the entire model. The resulting model is designed to handle a variety of scientific questions and instructions, making it a valuable tool for academic and research-oriented applications.

  - **Developed by:** ps2program
  - **Model type:** Causal Language Model (LoRA-finetuned)
  - **Language(s) (NLP):** English
  - **License:** Apache-2.0
  - **Finetuned from model:** `gemma3-270m`

### Model Sources

  - **Repository:** [https://huggingface.co/ps2program/gemma3-270m-scisinstruct](https://huggingface.co/ps2program/gemma3-270m-scisinstruct)
  - **Dataset Card:** [https://huggingface.co/datasets/zd21/SciInstruct](https://huggingface.co/datasets/zd21/SciInstruct)

## Uses

### Direct Use

This model is intended for direct use in applications requiring scientific text generation. Examples of its use include:

  - Answering questions about scientific principles.
  - Generating summaries of scientific papers or concepts.
  - Assisting in educational contexts for students and researchers.

### Downstream Use

This model can serve as a base for further fine-tuning on highly specialized, domain-specific scientific corpora (e.g., specific fields like biochemistry or astrophysics) to improve performance on those particular tasks.

### Out-of-Scope Use

The model is **not intended** for:

  - General-purpose conversation or casual chat.
  - Providing medical, legal, or financial advice.
  - Generating content in non-scientific domains where it may produce inaccurate or nonsensical outputs.
  - Applications where factual accuracy is critical without human verification.

## Bias, Risks, and Limitations

  - **Factual Inaccuracies:** The model may generate factually incorrect or outdated information. Users should always verify outputs, especially in academic or research contexts.
  - **Data Bias:** The model's performance and outputs are limited by the quality and content of its training data (`SciInstruct`). It may reflect any biases present in the original dataset.
  - **Limited Scope:** The model is specialized for scientific reasoning and may perform poorly on tasks outside of this domain.

### Recommendations

  - Always verify the generated scientific content with reliable sources.
  - Do not use the model for high-stakes decision-making.
  - Be aware of the model's limitations and potential for generating incorrect information.

## How to Get Started with the Model

To use this model, you'll need to load the base `gemma3-270m` model first, and then load the LoRA adapters on top of it.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load the base model (you must have access to gemma3-270m)
base_model_id = "google/gemma-3-270m"
model_id = "ps2program/gemma3-270m-scisinstruct"

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load the LoRA adapters from your repo
model = PeftModel.from_pretrained(base_model, model_id)

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Prepare a scientific prompt
prompt = "Explain the significance of CRISPR-Cas9 technology in genetic engineering."
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

# Generate the output
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.7,
    top_p=0.95
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Training Details

### Training Data

The model was fine-tuned on the `zd21/SciInstruct` dataset, which contains high-quality, scientifically-grounded instruction-following examples.

### Training Procedure

#### Preprocessing

The dataset was formatted into instruction-following prompts and tokenized using the `gemma3-270m` tokenizer.

#### Training Hyperparameters

  - **Training regime:** Mixed precision (bfloat16)
  - **LoRA parameters:** `r=8`, `lora_alpha=16`
  - **Optimizer:** `adamw_torch_fused`
  - **Learning rate:** `2e-5`
  - **Batch size:** `4`
  - **Epochs:** `5`
  - **Gradient accumulation steps:** `4`

## Evaluation

### Testing Data, Factors & Metrics

The model's performance was evaluated qualitatively by observing its ability to generate coherent and scientifically accurate responses to a diverse set of prompts. Standard metrics were not used in this initial fine-tuning.

### Results

The model demonstrates an improved ability to follow scientific instructions compared to the base `gemma3-270m` model. It can provide well-structured explanations and summaries.

## Environmental Impact

Carbon emissions were not calculated for this fine-tuning process. However, as a small-parameter model fine-tuned using LoRA, the training was computationally efficient and had a significantly lower environmental impact than training a large model from scratch.

  - **Hardware Type:** (e.g., 1x NVIDIA A100 GPU)
  - **Hours used:** (e.g., \~2 hours)
  - **Cloud Provider:** (e.g., Google Cloud, AWS, etc.)
  - **Compute Region:** (e.g., us-east-1)
  - **Carbon Emitted:** [More Information Needed]

## Citation

**BibTeX:**

```bibtex
@misc{ps2program2025gemma3,
  title={Gemma 3 270M SciInstruct LoRA},
  author={ps2program},
  year={2025},
  howpublished={\url{https://huggingface.co/ps2program/gemma3-270m-scisinstruct}},
  note={LoRA fine-tuning of Gemma 3 270M on the SciInstruct dataset.}
}
```

**APA:**

ps2program. (2025). *Gemma 3 270M SciInstruct LoRA*. Hugging Face Model Hub. [https://huggingface.co/ps2program/gemma3-270m-scisinstruct](https://huggingface.co/ps2program/gemma3-270m-scisinstruct)

## Model Card Authors

ps2program

## Model Card Contact

[Your preferred contact method or email address]