masterthesis-playground/raft

Retrieval-Augmented Finetuning (RAFT)

Voraussetzungen

• Generelles Preprocessing (Voraussetzung für BERTopic)
• BERTopic
  • Klassifikation muss durchgeführt sein, data/intermediate/culture_reviews.csv muss existieren

Vorbereiten des Retrieval-Corpus

python prepare_corpus.py --input_csv ../data/intermediate/culture_reviews.csv --out_dir out

Erstellen des RAFT-Datensatzes

python make_raft_data.py --out_dir out --n_examples 10

Training der QLoRA-Adapter

python train_mistral_raft.py --train_jsonl out/raft_train.jsonl --out_dir out/mistral_balitwin_lora

Inferenz

Per Baseline Mistral 7B + PEFT-Adapter

python rag_chat.py --lora_dir out/mistral_balitwin_lora

Pre-Merged Modell + Adapter

python rag_chat_merged.py --model_dir /path/to/model_folder --out_dir out