ptv3-lora-qwen7b

Userstory: From Voice Message to Psychotherapy Report End-to-end case study: Voice message -> transcription -> LoRA-generated psychotherapy report (PTV-3) -> PDF document. Built as a portfolio project in three stages on a single RTX 5090 (32 GB VRAM).

QLoRA adapter for generating structured German psychotherapy reports (PTV-3 / Bericht an den Gutachter) from therapy session transcripts.

Overview

Property	Value
Base model	Qwen2.5-7B-Instruct (4-bit)
Method	QLoRA via Unsloth
LoRA rank	16 (alpha 32)
Trainable params	40M / 7.6B (0.53%)
Adapter size	155 MB
Training	3 epochs, 280 samples, ~90 min on RTX 5090
Final loss	0.37

Training Details

• Dataset: John-creative/ptv3-bericht-lora-de-300 (280 train / 31 val)
• Framework: Unsloth 2026.4.4 + TRL 0.24.0 SFTTrainer
• Response-only training: Loss computed only on assistant JSON tokens via train_on_responses_only()
• Hardware: NVIDIA RTX 5090 (32 GB), bf16
• Optimizer: AdamW, cosine LR schedule, lr=2e-4, warmup 10 steps

Evaluation

Metric	Value
Parse rate (valid PTV3Report)	~60%
Verfahren exact match	~89%
Cosine similarity (vs teacher)	0.92
ROUGE-L konsiliarbericht	0.52
ROUGE-L befund	0.51
ICD exact match	~30%

Usage

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="John-creative/ptv3-lora-qwen7b",
    max_seq_length=4096,
    dtype=None,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "Du bist ein Assistenzsystem, das aus einer dokumentierten Psychotherapie-Sitzung einen strukturierten Bericht an den Gutachter nach PTV 3 als JSON erzeugt. Gib ausschließlich gültiges JSON zurück, ohne zusätzlichen Text."},
    {"role": "user", "content": "<session transcript here>"},
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(input_ids=inputs, max_new_tokens=2048, temperature=0.1)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Intended Use

Portfolio / educational project demonstrating QLoRA fine-tuning for structured clinical report generation. Not for clinical use.

Related

• Dataset: John-creative/ptv3-bericht-lora-de-300
• Code: John-Loschnig on GitHub