gemma-2-2b-it-ko

Gemma-2-2b-it Fine-Tuned on KoAlpaca-v1.1a

<!-- Provide a quick summary of what the model is/does. -->

Model Description

<!-- Provide a longer summary of what this model is. -->

This model is a fine-tuned version of the google/gemma-2-2b-it model on the Korean dataset beomi/KoAlpaca-v1.1a. It is designed to generate coherent, contextually appropriate responses in Korean. The fine-tuning process has enhanced the model's ability to handle conversational prompts in a colloquial style, responding with contextually aware and polite expressions.

The base model, Gemma-2-2B-it, is a large pre-trained language model built for multilingual text generation tasks. With the fine-tuning on the KoAlpaca dataset, the model has been optimized to perform better on Korean text generation, offering more natural and conversational outputs.

Training Process

<!-- Provide the basic links for the model. -->

The model was fine-tuned using the KoAlpaca-v1.1a dataset, which is designed for instruction-following tasks in Korean. The dataset contains various examples of questions and corresponding responses in Korean, which helped the model learn polite conversational structures.

Dataset

• Dataset Used: beomi/KoAlpaca-v1.1a
• Type of Data: Instruction-following examples in Korean, with both the instruction and expected response provided in each entry. Training Configuration
• Model Base: google/gemma-2-2b-it
• LoRA Configuration: Applied LoRA with the following parameters:
• Quantization: 4-bit quantization (bnb_4bit) for efficient fine-tuning
• Training Hyperparameters:
  • Steps: 3000
  • Learning Rate: 2e-4
  • Batch Size: 1
  • Warmup Steps: 100
  • Gradient Accumulation: 4 steps
  • Optimizer: paged_adamw_8bit
  • Precision: FP16

Results and Performance

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

Example Input: “배가 고파서 마라탕을 먹었어요.” (I was feeling hungry, so I ate maratang.)

• Before Fine-tuning: Output:“마라탕 (maratang): This is a Korean soup made with various ingredients like meat, vegetables, and noodles.
  고파 (gopa): This means ‘to be hungry.’
  먹었어요 (meok-eosseoyo): This is the polite way to say ‘I ate.’”

• After Fine-tuning: Output: “맛있게 드셨군요! 저도 이렇게 하면 좋겠습니다. 내일은 어떤 음식을 해볼까 생각해보세요?” (It sounds like you enjoyed your meal! I should try that too. What do you plan to cook tomorrow?)

The fine-tuned model shows a significant improvement in contextual understanding and produces more conversational and polite responses in Korean. It also demonstrates an ability to provide helpful follow-up suggestions, which is essential in conversational agents.

Future Work

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

• Further fine-tuning on larger or more diverse Korean datasets could improve the model's versatility.
• Exploring different LoRA configurations and quantization techniques could yield more efficient results for deployment on smaller devices.
• Evaluation with human raters to measure improvements in conversation quality.