Qwen3.5 0.8B NTKT Vietnamese Conversation SFT
This PEFT LoRA adapter fine-tunes Qwen/Qwen3.5-0.8B for conversational knowledge tracing. It predicts Correct or Incorrect from a learner's Vietnamese interaction history represented as full text.
Training
- Method: supervised fine-tuning with assistant-only loss
- Context length: 15,000 tokens
- Steps: 100
- Effective batch size: 8
- LoRA rank/alpha: 16/16
- Trainable parameters: 6,389,760 (0.74%)
- Final training loss: 0.3562
View the training run on Weights & Biases.
Evaluation
Evaluation used a learner-disjoint validation split with 115,853 labeled predictions across 2,361 conversations.
| Metric | Value |
|---|---|
| Accuracy | 0.7326 |
| F1 | 0.8287 |
| Precision | 0.7580 |
| Recall | 0.9139 |
| ROC-AUC | 0.7320 |
| Brier score | 0.1784 |
| Log loss | 0.5323 |
Usage
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer
repo_id = "bihungba1101/ntkt-conversation-vietnamese-full-text-qwen3.5-0.8b-sft"
model = AutoPeftModelForCausalLM.from_pretrained(repo_id, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(repo_id)
The adapter expects the included chat template and the same full-text conversation representation used during training.
Framework versions
- PEFT 0.18.1
- TRL 0.23.1
- Transformers 5.2.0
- PyTorch 2.10.0+cu128
- Datasets 4.3.0
- Tokenizers 0.22.2
- Downloads last month
- 18