dtp-fine-tuning/multi-turn_chatbot_diploy

Model Card for SFT-Bakti-8B-Base-MultiTurn-Chatbot

Model Details

Model Description

This model is a fine-tuned version of [aitfindonesia/Bakti-8B-Base] designed specifically for multi-turn conversational capabilities in the Indonesian language. It was trained using the Unsloth library for faster and memory-efficient training, utilizing LoRA (Low-Rank Adaptation).

The model is optimized to handle context retention across multiple turns of conversation, making it suitable for interview simulations, customer support, and general-purpose Indonesian assistants.

• Developed by: DTP Fine Tuning Team
• Model type: Causal Language Model (Fine-tuned Qwen2/3 architecture)
• Language(s) (NLP): Indonesian
• License: Apache 2.0
• Finetuned from model: aitfindonesia/Bakti-8B-Base

Uses

Direct Use

The model is designed for:

• Multi-turn chat interactions in Indonesian.
• Question Answering (QA) requiring context from previous turns.
• Roleplay interactions (e.g., interview scenarios).

Out-of-Scope Use

• The model should not be used for generating factually accurate data without RAG (Retrieval Augmented Generation) as hallucinations are possible.
• Not intended for code generation tasks.

Training Details

Training Data

Dataset: dtp-fine-tuning/dtp-multiturn-interview-valid-15k

• Split: Train (90%) / Test (10%)
• Format: Multi-turn conversation format.
• Max Length: 2048 tokens

Training Procedure

The model was fine-tuned using Unsloth on a single NVIDIA A100 (80GB) GPU. It utilizes 4-bit quantization (NF4) to reduce memory usage while maintaining performance via QLoRA.

Training Hyperparameters

• Training regime: QLoRA (4-bit quantization with FP16 precision)
• Optimizer: AdamW 8-bit
• Learning Rate: $2 \times 10^{-5}$
• Scheduler: Linear with 5% warmup
• Batch Size: 8 per device (Gradient Accumulation: 4)
• Epochs: 2
• LoRA Config:
  • Rank ($r$): 16
  • Alpha ($\alpha$): 32
  • Dropout: 0.05
  • Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Hardware

• GPU: NVIDIA A100 80GB PCIe
• VRAM Usage: Peak allocation approx. 19GB (23% utilization) due to 4-bit loading.

Evaluation

Results

The model demonstrates strong convergence on the multi-turn dataset.

• Final Train Loss: $\approx 0.42$
• Final Eval Loss: $\approx 0.41$

Note: The model outperforms the standard Qwen3-8B baseline on this specific Indonesian dataset, achieving lower loss values faster.

Environmental Impact

• Hardware Type: NVIDIA A100 80GB
• Compute Region: asia-east1
• Carbon Emitted: 0.31

Framework Versions

• Unsloth
• PEFT
• Transformers
• TRL