roc-hci/HAL-Qwen14B

HAL (Human Aligning LLMs)

HAL is an alignment framework that makes LLMs more human-like in conversation by optimizing for interpretable conversational traits derived from contrastive dialogue data. It uses a transparent reward signal to align models via preference optimization, improving perceived human-likeness without degrading core capabilities.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

model_path = "roc-hci/HAL-Qwen14B-8bit"

bnb_config = BitsAndBytesConfig(
    load_in_8bit=True
)

tokenizer = AutoTokenizer.from_pretrained(model_path)

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    quantization_config=bnb_config,
    device_map="auto"
)

model.eval()

prompt = """Generate a full conversation of the following person at the doctor's visit.
Give them a unique personality based on their biography. Follow a linguistic style suitable for the person. Vary the statement lengths to make it natural.
Don't say anything else.

<input_data>
Biography: Elliot Meyer is a 25-year-old researcher based in Boston. Known for his deep, contemplative demeanor, Elliot was always that child who asked too many questi$

Medical condition: chronic tension headaches

Reason for clinical visit: Elliot is visiting his primary care physician to discuss worsening tension headaches that have started to interfere with his work and concen$
</input_data>

<dialogue_format>
P: ...
D: ...
P: ...
D: ...
</dialogue_format>

<dialogue>"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=100,
        do_sample=True,
        temperature=0.7
    )

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