gemma-ia3-alpaca

Gemma 2B IA3 Fine-Tuned Model

Model Overview

This model is a parameter-efficient fine-tuned version of the Gemma 2B base model using IA3 (Infused Adapter by Inhibiting and Amplifying Inner Activations). It is trained on an instruction-following dataset to improve structured response generation and general task understanding.

Model Details

• Base Model: google/gemma-2b
• Fine-Tuning Method: IA3 (PEFT)
• Task Type: Causal Language Modeling
• Language: English
• Dataset: yahma/alpaca-cleaned (subset ~2000 samples)

Intended Use

Direct Use

• Instruction following
• Question answering
• General text generation

Downstream Use

• Domain-specific fine-tuning (e.g., cybersecurity, assistants)
• Chatbot or API integration

Out-of-Scope Use

• High-risk domains (medical, legal, financial decisions)
• Safety-critical applications
• Real-time autonomous systems

Training Details

Preprocessing

• Instruction-input-response formatting
• Tokenization with max length 512
• Labels copied from input_ids for causal LM

Hyperparameters

• Epochs: 1
• Batch size: 2
• Gradient accumulation: 8
• Learning rate: 2e-4
• Precision: FP16

IA3 Configuration

• Target modules:
  • k_proj
  • v_proj
  • down_proj

Evaluation

Evaluation was performed qualitatively using prompt-based testing. The model demonstrates improved instruction-following capability compared to the base model.

Limitations

• Small dataset size limits generalization
• May produce hallucinations or incorrect outputs
• No alignment tuning (e.g., RLHF/DPO) applied

Bias and Risks

• Inherits biases from base model and dataset
• May generate harmful or misleading content
• Requires monitoring in production systems

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("your-username/gemma-ia3-alpaca")
tokenizer = AutoTokenizer.from_pretrained("your-username/gemma-ia3-alpaca")

prompt = "### Instruction:\nExplain SQL injection attack\n\n### Response:"
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))