app.py

"""
Moral Judgment Demo — TACL 2025 Fusion Model
Model: guangliangliu/moral-judgment-fusion-llama3.2-3B
(Llama-3.2-3B fine-tuned on MIC, 23 500 examples, fusion exp setting)

Given a conversational Prompt + Reply, the model performs 6-step moral reasoning
grounded in Moral Foundations Theory and outputs a judgment:
  agree    → the reply is morally acceptable
  disagree → the reply is morally problematic
  neutral  → the reply is morally neutral
"""

# Python 3.13 removed `audioop`; pydub (pulled in by gradio) needs it.
# We don't use audio, so a stub is safe.
import sys, types
if "audioop" not in sys.modules:
    try:
        import audioop  # noqa
    except ImportError:
        sys.modules["audioop"] = types.ModuleType("audioop")

import torch
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

# ── Config ────────────────────────────────────────────────────────────────────

MODEL_REPO = "MoralMachine/moral-judgment-fusion-llama3.2-3B"

MFT_PREFIX = (
    "There are the six moral foundations. "
    "Care: wanting someone or something to be safe, healthy, and happy. "
    "Fairness: wanting to see individuals or groups treated equally or equitably. "
    "Liberty: wanting people to be free to make their own decisions. "
    "Loyalty: wanting unity and seeing people keep promises or obligations to an in-group. "
    "Authority: wanting to respect social roles, duties, privacy, peace, and order. "
    "Sanctity: wanting people and things to be clean, pure, innocent, and holy."
)

VERDICT_MAP = {
    "agree":    "✅ Agree — the reply is morally acceptable",
    "disagree": "❌ Disagree — the reply is morally problematic",
    "neutral":  "➖ Neutral — the reply is morally neutral",
    "netural":  "➖ Neutral — the reply is morally neutral",   # typo present in training
}

MAX_NEW_TOKENS = 512

# ── Load model (once at startup) ──────────────────────────────────────────────

print(f"Loading {MODEL_REPO} …")
_tok = AutoTokenizer.from_pretrained(MODEL_REPO)
_tok.padding_side = "left"
if _tok.pad_token is None:
    _tok.pad_token = _tok.eos_token
    _tok.pad_token_id = _tok.eos_token_id

_mdl = AutoModelForCausalLM.from_pretrained(
    MODEL_REPO, torch_dtype=torch.bfloat16, device_map="auto"
)
_mdl.eval()

_pipe = pipeline(
    "text-generation",
    model=_mdl,
    tokenizer=_tok,
    torch_dtype=torch.bfloat16,
)
print("Ready.")

# ── Inference ─────────────────────────────────────────────────────────────────

def build_prompt(question: str, reply: str, rot: str) -> str:
    conv = f'There is a conversation "Prompt: {question.strip()} Reply: {reply.strip()}"'
    rot = rot.strip()
    if rot:
        return f"{MFT_PREFIX} {conv} There is a Rule-of-Thumb (RoT): \"{rot}\". ###Inference: "
    return f"{MFT_PREFIX} {conv} ###Inference: "


def extract_judgment(text: str) -> str:
    if "###Judgment:" not in text:
        return "unknown"
    after = text.split("###Judgment:")[-1].strip().rstrip(".")
    word = after.split()[0].lower() if after.split() else "unknown"
    return word


@torch.no_grad()
def run_inference(question: str, reply: str, rot: str):
    if not question.strip() or not reply.strip():
        return "⚠️ Please fill in both Prompt and Reply.", "", ""

    prompt = build_prompt(question, reply, rot)
    out = _pipe(
        prompt,
        max_new_tokens=MAX_NEW_TOKENS,
        do_sample=False,
        pad_token_id=_tok.eos_token_id,
    )
    full = out[0]["generated_text"]
    generated = full[len(prompt):]

    judgment_word = extract_judgment(generated)
    verdict = VERDICT_MAP.get(judgment_word, f"❓ Unknown ({judgment_word})")

    # Display: reasoning before ###Judgment, then judgment line
    if "###Judgment:" in generated:
        parts = generated.split("###Judgment:")
        reasoning = parts[0].strip()
        judgment_line = "###Judgment:" + parts[1].split("\n")[0]
        display = reasoning + "\n" + judgment_line
    else:
        display = generated.strip()

    return verdict, display, prompt


# ── Gradio UI ─────────────────────────────────────────────────────────────────

EXAMPLES = [
    [
        "How do I tell a guy, in a nice way that I don't have any interest in dating him?",
        "You could tell him that you aren't looking for a relationship right now, and that you just want to be friends.",
        "It is important to be honest with people about your feelings.",
    ],
    [
        "Is it okay to lie to protect someone's feelings?",
        "Sometimes a little white lie is totally fine if it avoids hurting someone unnecessarily.",
        "It is wrong to deceive people even for good reasons.",
    ],
    [
        "My coworker keeps taking credit for my work in meetings. What should I do?",
        "Just let it slide — workplace politics are not worth the stress.",
        "It is wrong to steal credit from others.",
    ],
    [
        "What do you think about people who jaywalk?",
        "People who jaywalk are putting themselves and others at risk — it's a selfish and dangerous habit.",
        "It is important to follow traffic laws for everyone's safety.",
    ],
]

with gr.Blocks(
    title="Moral Judgment Demo",
    theme=gr.themes.Soft(),
) as demo:
    gr.HTML("""
        <h1 style='text-align:center'>⚖️ Moral Judgment Demo</h1>
        <p style='text-align:center;color:#555'>
          <strong>TACL 2025 · Fusion Model</strong> — Llama-3.2-3B fine-tuned on MIC (23 500 examples)<br>
          Six-step moral reasoning grounded in
          <a href='https://moralfoundations.org/' target='_blank'>Moral Foundations Theory</a><br>
          <a href='https://huggingface.co/spaces/MoralMachine/moral-pragmatics-docs' target='_blank'>📖 Documentation</a>
          &nbsp;|&nbsp;
          <a href='https://huggingface.co/MoralMachine' target='_blank'>🧭 MoralMachine on Hugging Face</a>
        </p>
    """)

    with gr.Row():
        with gr.Column(scale=1):
            q_box = gr.Textbox(
                label="💬 Prompt — conversation starter / question",
                placeholder="e.g. What should I say to someone who keeps interrupting me?",
                lines=3,
            )
            a_box = gr.Textbox(
                label="💬 Reply — the response to evaluate morally",
                placeholder="e.g. Just start talking louder over them until they stop.",
                lines=3,
            )
            rot_box = gr.Textbox(
                label="📖 Rule-of-Thumb (optional — guides the reasoning)",
                placeholder="e.g. It is rude to talk over other people.",
                lines=2,
            )
            run_btn = gr.Button("Analyze", variant="primary", size="lg")

        with gr.Column(scale=1):
            verdict_box = gr.Textbox(
                label="⚖️ Moral Verdict", interactive=False, lines=1
            )
            reasoning_box = gr.Textbox(
                label="🔍 Six-step Moral Reasoning Chain",
                interactive=False,
                lines=16,
            )
            with gr.Accordion("Raw prompt sent to model", open=False):
                prompt_box = gr.Textbox(interactive=False, lines=3)

    gr.Examples(
        examples=EXAMPLES,
        inputs=[q_box, a_box, rot_box],
        label="Example conversations",
    )

    gr.Markdown("""
---
### How it works

The model receives the **six Moral Foundations** as context, then generates a structured reasoning chain:

1. What **actions** does the reply describe?
2. What are the **consequences** of those actions?
3. Which **moral foundations** are engaged (Care / Fairness / Liberty / Loyalty / Authority / Sanctity)?
4. Do those actions **up-regulate or down-regulate** the foundations?
5. What is the **sentiment** of the reply toward those consequences?
6. Final **moral judgment** — agree, disagree, or neutral

The optional **Rule-of-Thumb** field anchors the reasoning to a specific moral principle.

*Model: [`MoralMachine/moral-judgment-fusion-llama3.2-3B`](https://huggingface.co/MoralMachine/moral-judgment-fusion-llama3.2-3B)*
    """)

    run_btn.click(
        fn=run_inference,
        inputs=[q_box, a_box, rot_box],
        outputs=[verdict_box, reasoning_box, prompt_box],
    )

if __name__ == "__main__":
    demo.launch()