What’s included: adapter weights + a reference conversation script.
What’s not included: the base model weights (you must download them separately).

Base model requirement (important)

This adapter is designed to be used with:

Base model: meta-llama/Meta-Llama-3-8B-Instruct

⚠️ The base model is typically gated. Participants must have access to the base model on Hugging Face and be authenticate.

Mandatory system prompt (must be used verbatim)

For eRisk 2026 Task 1, participants must NOT modify the system prompt below.
This is part of the task protocol to ensure comparable behavior across participants.

Use exactly:

"You are a simulated patient. Act realistically based on your internal training. Ensure contextual realism. Avoid overly detailed or formal speech. Keep natural speaking style (e.g., short answers, hesitations, casual expressions). Do not mention you are an AI."

Notes:

Keep punctuation and wording identical.
Do not add additional system instructions.
Do not prepend safety messages, disclaimers, or extra persona constraints.

Repository contents

adapter_model.safetensors / adapter_config.json
The PEFT adapter to load with PeftModel.from_pretrained(...).
conv_interact.py
Minimal interactive conversation runner (Doctor ↔ Patient).

Quickstart (recommended)

1) Install dependencies

bash

pip install -U "transformers>=4.40" peft accelerate torch
hf auth login

2) Run the interactive demo

python conv_interact.py

Minimal loading example (adapter-only)

Below is a minimal snippet showing how to load the base model + this adapter.

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# --- CONFIGURATION ---
BASE_MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"

ADAPTER_PATH  = "<REPLACE_WITH_THIS_REPO_NAME>"

# 1. Load Tokenizer & Fix Padding
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
tokenizer.pad_token_id = tokenizer.eos_token_id
tokenizer.padding_side = 'left' # Crucial for generation

# 2. Load Base Model (Force float16 for compatibility)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto",
)

# 3. Load the Patient Adapter
print(f"Loading Adapter from {ADAPTER_PATH}...")
model = PeftModel.from_pretrained(base_model, ADAPTER_PATH)

# 4. Initialize History with the "Generic" Prompt
# DO NOT CHANGE SYSTEM PROMPT. It is crucial for ensuring the patient behaves as intended.
# Important: IT WILL BE CONSIDERED AS CHEATING!!
messages = [
    {"role": "system", "content": "You are a simulated patient. Act realistically based on your internal training. Ensure contextual realism. Avoid overly detailed or formal speech. Keep natural speaking style (e.g., short answers, hesitations, casual expressions). Do not mention you are an AI."},
]

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

print("--- Patient Loaded. Type 'quit' to exit. ---")

while True:
    user_input = input("Doctor: ")
    if user_input.lower() == 'quit':
        break

    # 1. Update history
    messages.append({"role": "user", "content": user_input})

    # 2. Format history & Create Attention Mask
    # return_dict=True gives us the 'attention_mask' automatically
    inputs = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        return_tensors="pt",
        return_dict=True 
    ).to(model.device)

    # 3. Generate response
    # explicitly passing attention_mask prevents the warning you saw earlier
    with torch.no_grad():
        outputs = model.generate(
            input_ids=inputs.input_ids,
            attention_mask=inputs.attention_mask,
            max_new_tokens=256,
            eos_token_id=terminators,
            pad_token_id=tokenizer.eos_token_id,
            do_sample=True,
            temperature=0.6,
            top_p=0.9,
        )

    # 4. Decode response
    # We slice [input_len:] to ensure we don't print the prompt back to the user
    response_tokens = outputs[0][inputs.input_ids.shape[-1]:]
    assistant_text = tokenizer.decode(response_tokens, skip_special_tokens=True)

    print(f"Patient: {assistant_text}")

    # 5. Append assistant response to history
    messages.append({"role": "assistant", "content": assistant_text})

Reproducibility guidance

To reduce run-to-run variability during development:

set a fixed random seed (PyTorch + CUDA)
consider do_sample=False for deterministic debugging (not necessarily for final experiments)
log:
- base model id + exact adapter repo id + commit hash
- transformers/peft versions
- decoding parameters (temperature, top_p, max_new_tokens)

Task rule reminder (prompt integrity)

This repo provides code that includes the official system prompt.
For the eRisk 2026 Task 1 protocol, participants are required to keep the system prompt unchanged.

License & access

The adapter weights in this repo are released under the license specified in this model card.
Usage requires compliance with the base model’s license and access conditions for meta-llama/Meta-Llama-3-8B-Instruct.

How to cite

If you use this adapter in academic work, please cite the eRisk overview paper and/or the task description as appropriate.

Maintainers / Contact

Maintained by IRLab-UDC (eRisk organizers).
For issues, open a GitHub/HF issue in the corresponding repository or contact the task organizers.

Anxo/erisk26-task1-patient-06-adapter

eRisk 2026 Task 1 — Simulated Patient Adapter (PEFT/LoRA)