klemenp950/gams-email-finetuned-v2

Model Card for klemenp950/gams-email-finetuned

Fine-tuned version of GaMS-1B for generating formal Slovenian patient-style emails from appointment-setting conversations.

---

Model Details

Model Description

This model is a LoRA fine-tuned variant of the Slovenian GaMS-1B causal language model. It is trained to generate structured, polite, and formal emails based on conversational transcripts involving appointment scheduling and basic medical administrative communication.

The model receives a conversation transcript followed by a special <sep> token and generates the corresponding email.

• Developed by: Klemen P.
• Model type: Causal Language Model (decoder-only Transformer)
• Language(s): Slovenian
• License: Same as base model (GaMS-1B)
• Finetuned from model: cjvt/GaMS-1B

Model Sources

• Base model: https://huggingface.co/cjvt/GaMS-1B
• Fine-tuned model: https://huggingface.co/klemenp950/gams-email-finetuned

---

Uses

Direct Use

• Generate formal Slovenian emails from appointment-related conversations
• Assist healthcare administrative staff with drafting emails
• Prototype conversational-to-email generation systems

Downstream Use

• Further fine-tuning for other administrative domains
• Customer support email generation
• Domain-adapted Slovenian text generation

Out-of-Scope Use

• Medical diagnosis
• Legal advice
• Emergency or safety-critical communication
• Autonomous decision-making without human review

---

Bias, Risks, and Limitations

• Trained on a limited and domain-specific dataset
• May hallucinate information
• May reflect biases present in training data
• Not medically validated
• Sensitive to prompt formatting

Recommendations

• Keep a human in the loop
• Review generated emails before sending
• Do not rely on outputs for medical decision-making

---

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "klemenp950/gams-email-finetuned"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.float16
)

conversation = """spk_0: Pozdravljeni, kaj je razlog vašega obiska?
spk_1: povišan krvni tlak
spk_0: Kako vam je ime?
spk_1: Zdenka Novak
spk_0: Katerega zdravnika bi radi obiskali?
spk_1: Dr. Kos
spk_0: Kdaj bi radi prišli?
spk_1: torek, 24. februarja ob 10:00
spk_0: Najlepša hvala! Veselimo se vašega srečanja. Lep dan!"""

prompt = f"<s>{conversation}<sep>"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))