ksumanthreddy96/studymate-ctp-sft

StudyMate CTP — SFT Adapter StudyMate is an experimental domain-adapted language-model project for question answering about Construction Technology and Practices (CTP) lecture material. This repository contains the Supervised Fine-Tuning (SFT) LoRA adapter trained on top of the StudyMate DAPT adapter. It is an adapter, not a standalone full model. Model lineage

• Original model family: Qwen2.5-0.5B
• DAPT adapter: ksumanthreddy96/studymate-ctp-dapt
• This repository: SFT adapter fine-tuned from the DAPT adapter
• Fine-tuning approach: QLoRA-style training with 4-bit model loading and LoRA adapters
• Training tools: Unsloth, Hugging Face Transformers, and PEFT Training data The source material consisted of 11 scanned CTP lecture PDFs. Text was extracted and preprocessed using Python, PyMuPDF, Tesseract OCR, and EasyOCR. Instruction-style question-and-answer examples were then prepared from the extracted lecture content.
• SFT training examples: 500
• SFT evaluation examples: 150
• Epochs configured: 3
• Learning rate: 1e-5
• Maximum sequence length: 256 The source PDFs were also used during DAPT, so the SFT evaluation set is held out from SFT training but is not an independent test on unseen source documents. Evaluation results The following results were recorded for the SFT evaluation run: Metric Result Evaluation loss 3.7138 Perplexity 41.0083 BLEU-1 0.1344 BLEU-2 0.0944 BLEU-3 0.0738 BLEU-4 0.0599 Mean token F1 0.1852 Mean ROUGE-L 0.1777 Exact-match rate 0%

BLEU and ROUGE-L measure overlap with reference answers; they do not directly measure factual correctness. The scores indicate that generated answers still need improvement and human review. Intended use

• Experimentation with domain adaptation and parameter-efficient fine-tuning.
• Educational prototyping for CTP lecture question answering.
• Demonstrating a DAPT → SFT workflow for a small language model. Limitations
• OCR-derived text can contain mistakes, missing words, and formatting artifacts.
• Generated answers may be incomplete, repetitive, off-topic, or factually incorrect.
• Evaluation metrics are text-overlap metrics and do not establish factual accuracy.
• This repository contains an adapter. Use it with its compatible base-model lineage and tokenizer/configuration rather than treating it as a complete standalone model.
• Do not use the model as the sole source for engineering, safety, or other high-stakes decisions. Loading the adapter The adapter was trained on top of the StudyMate DAPT adapter. Use a compatible Unsloth/Transformers/PEFT environment and load the SFT adapter with its expected base model. Check the training notebook for the exact loading and tokenizer setup before running inference. Data and licensing The training material came from lecture PDFs. The original PDFs and derived training examples are not included here by default. Ensure you have permission to redistribute source material or derived data before sharing them. No license is specified in this model card because redistribution permissions for all underlying material have not been established.