pelosi70/jch1

Model Summary

This model is a Korean instruction-following Small Language Model (SLM) fine-tuned from the Llama-3.2-3B base model using Supervised Fine-Tuning (SFT). The objective of this model is to validate a resource-efficient fine-tuning and deployment pipeline suitable for on-premise and constrained GPU/CPU environments, rather than to maximize benchmark scores.

---

Training Approach

• Base Model: Meta Llama-3.2-3B (base, non-instruct)
• Fine-Tuning Method: Supervised Fine-Tuning (SFT)
• Parameter-Efficient Training: LoRA (PEFT)
• Quantization During Training: 4-bit (QLoRA)
• Training Framework: Unsloth + Hugging Face TRL
• Training Environment: Single GPU (Google Colab, Tesla T4)

The model was trained using an instruction–response prompt template (Alpaca-style), enabling stable instruction-following behavior in Korean. The fine-tuning process focused on maintaining the base model’s general language capability while adapting response style, tone, and instruction compliance.

---

Dataset

• Primary Dataset: korean_safe_conversation
• Language: Korean
• Data Type: Instruction–response conversational data
• Data Scale: ~27K samples

The dataset was preprocessed to ensure:

• Clear separation between instruction and response
• Explicit end-of-sequence (EOS) control to prevent uncontrolled generation
• Consistent prompt formatting for stable training behavior

---

Intended Use

This model is intended for:

• Korean instruction-following assistants

• Domain-adapted SLM experimentation

• On-premise inference scenarios where:

  • Data privacy is critical
  • GPU resources are limited
  • Low-latency local inference is preferred

Typical application examples include:

• Internal enterprise assistants
• Document-based Q&A systems (pre/post-RAG)
• Operational report generation from structured or semi-structured text

---

Deployment

• Format: GGUF
• Quantization: Q8
• Deployment Target: CPU or low-VRAM environments
• Distribution: Hugging Face Hub

The GGUF format allows the model to be deployed without external API dependencies, making it suitable for secure, offline, or air-gapped environments.

---

Limitations

• This model is not an official Meta Instruct model
• Preference optimization methods such as DPO or RLHF were not applied
• The model was trained for behavior adaptation and stability, not for benchmark optimization
• Performance may vary outside the instruction-following and conversational domains

---

Technical Motivation

This project demonstrates that domain-adapted instruction-following models can be efficiently built and deployed using small-scale resources, providing a practical alternative to large, cost-intensive LLM deployments in real-world systems.