Create MIKE v4 Indian legal triage demo

README.md

@@ -1,13 +1,33 @@
 ---
-title: MIKE Demo
-emoji: 👀
-colorFrom: red
-colorTo: purple
+title: MIKE Indian Legal Triage
+emoji: "⚖️"
+colorFrom: indigo
+colorTo: blue
 sdk: gradio
-sdk_version: 6.19.0
-python_version: '3.13'
+sdk_version: 5.49.1
 app_file: app.py
 pinned: false
+license: llama4
+suggested_hardware: a100-large
+suggested_storage: medium
+models:
+- itsalloverig/MIKE
+datasets:
+- itsalloverig/adaption-indian-legal-triage-samples-v4
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# MIKE Indian Legal Triage Demo
+
+Interactive demonstration of the selected MIKE v4 controlled-fused LoRA
+adapter. MIKE supports India-focused legal research triage and is not a
+substitute for advice from a qualified legal professional.
+
+- Model: https://huggingface.co/itsalloverig/MIKE
+- Dataset: https://huggingface.co/datasets/itsalloverig/adaption-indian-legal-triage-samples-v4
+- Adaption held-out pairwise score: 92.16%
+- Separate Indian-law domain score: 65.66%
+
+The Space requires GPU hardware and persistent storage because the Llama 4
+Scout base is a large mixture-of-experts model. It deliberately does not fall
+back to a different model when the V4 checkpoint cannot be loaded.
+

app.py

@@ -0,0 +1,198 @@
+import os
+import threading
+from typing import Any
+
+import gradio as gr
+
+
+MODEL_ID = "itsalloverig/MIKE"
+BASE_MODEL_ID = "bnb-community/Llama-4-Scout-17B-16E-Instruct-bnb-4bit"
+DATASET_ID = "itsalloverig/adaption-indian-legal-triage-samples-v4"
+
+SYSTEM_PROMPT = """You are MIKE, an India-only English legal research-triage
+assistant. Preserve source-bounded facts. Never invent statutes, section or
+article numbers, cases, deadlines, forums, or outcomes. Cite a numbered
+authority only when it is present in the supplied context. Cover both legacy
+IPC, CrPC and Indian Evidence Act terminology and current BNS, BNSS and BSA
+terminology when the source supports it. Separate facts, assumptions, missing
+evidence, forum or remedy options, urgency, and next research steps. Frame all
+outputs as legal research assistance, not legal advice."""
+
+model: Any = None
+processor: Any = None
+load_error: str | None = None
+load_state = "waiting_for_gpu"
+load_lock = threading.Lock()
+
+
+def gpu_available() -> bool:
+    try:
+        import torch
+
+        return torch.cuda.is_available()
+    except Exception:
+        return False
+
+
+def load_model() -> None:
+    global model, processor, load_error, load_state
+    with load_lock:
+        if model is not None or load_state == "loading":
+            return
+        if not gpu_available():
+            load_state = "waiting_for_gpu"
+            return
+        try:
+            load_state = "loading"
+            import torch
+            from peft import PeftModel
+            from transformers import AutoProcessor, Llama4ForConditionalGeneration
+
+            processor = AutoProcessor.from_pretrained(MODEL_ID)
+            base = Llama4ForConditionalGeneration.from_pretrained(
+                BASE_MODEL_ID,
+                device_map="auto",
+                torch_dtype=torch.bfloat16,
+                low_cpu_mem_usage=True,
+            )
+            model = PeftModel.from_pretrained(base, MODEL_ID)
+            model.eval()
+            load_state = "ready"
+        except Exception as error:
+            load_error = f"{type(error).__name__}: {error}"
+            load_state = "failed"
+
+
+def status_markdown() -> str:
+    labels = {
+        "waiting_for_gpu": "Waiting for GPU hardware",
+        "loading": "Loading Llama 4 Scout and the MIKE v4 adapter",
+        "ready": "MIKE v4 is ready",
+        "failed": "Model loading failed",
+    }
+    result = f"**Runtime:** {labels.get(load_state, load_state)}"
+    if load_error:
+        result += f"\n\n`{load_error}`"
+    return result
+
+
+def generate_reply(
+    message: str,
+    history: list[dict[str, str]],
+    source_context: str,
+    max_new_tokens: int,
+) -> str:
+    if model is None:
+        load_model()
+    if model is None:
+        return (
+            "MIKE v4 is not loaded. This Space requires the configured GPU "
+            "hardware and persistent storage. No substitute model was used."
+        )
+
+    import torch
+
+    messages: list[dict[str, str]] = [{"role": "system", "content": SYSTEM_PROMPT}]
+    for item in history[-6:]:
+        role = item.get("role")
+        content = item.get("content")
+        if role in {"user", "assistant"} and isinstance(content, str):
+            messages.append({"role": role, "content": content})
+
+    user_content = message.strip()
+    if source_context.strip():
+        user_content += (
+            "\n\nSUPPLIED CONTEXT (cite only authorities explicitly present here):\n"
+            + source_context.strip()
+        )
+    messages.append({"role": "user", "content": user_content})
+
+    inputs = processor.apply_chat_template(
+        messages,
+        add_generation_prompt=True,
+        tokenize=True,
+        return_tensors="pt",
+        return_dict=True,
+    )
+    first_device = next(model.parameters()).device
+    inputs = {key: value.to(first_device) for key, value in inputs.items()}
+
+    with torch.inference_mode():
+        output = model.generate(
+            **inputs,
+            max_new_tokens=int(max_new_tokens),
+            do_sample=False,
+            repetition_penalty=1.05,
+        )
+    generated = output[0][inputs["input_ids"].shape[-1] :]
+    return processor.decode(generated, skip_special_tokens=True).strip()
+
+
+if gpu_available():
+    threading.Thread(target=load_model, daemon=True).start()
+
+
+CSS = """
+.gradio-container {max-width: 1100px !important;}
+.legal-note {border-left: 4px solid #4f46e5; padding-left: 12px;}
+"""
+
+with gr.Blocks(css=CSS, title="MIKE Indian Legal Triage") as demo:
+    gr.Markdown(
+        """
+        # ⚖️ MIKE — Indian Legal Research Triage
+
+        Ask an India-focused legal research question. Add authoritative source
+        text when you want citation-aware analysis.
+
+        <div class="legal-note"><strong>Important:</strong> MIKE provides legal
+        research assistance, not legal advice. Verify outputs against current
+        official sources and a qualified Indian legal professional.</div>
+        """
+    )
+    runtime = gr.Markdown(value=status_markdown, every=10)
+    context = gr.Textbox(
+        label="Optional supplied legal context",
+        placeholder="Paste the relevant statute, order, contract excerpt, or case material.",
+        lines=6,
+    )
+    token_limit = gr.Slider(
+        minimum=128,
+        maximum=1024,
+        value=640,
+        step=64,
+        label="Maximum response tokens",
+    )
+    gr.ChatInterface(
+        fn=generate_reply,
+        additional_inputs=[context, token_limit],
+        chatbot=gr.Chatbot(height=520, type="messages"),
+        textbox=gr.Textbox(
+            placeholder="Describe the legal issue, relevant dates, documents, and desired research outcome.",
+            lines=3,
+        ),
+        examples=[
+            [
+                "A complaint concerns conduct spanning June and August 2024. Explain what facts determine whether IPC/CrPC or BNS/BNSS terminology applies.",
+                "",
+                640,
+            ],
+            [
+                "Create a source-bounded document and evidence checklist for an employment termination dispute. Do not invent statutory sections.",
+                "",
+                640,
+            ],
+            [
+                "Triage a consumer dispute involving a defective online purchase and identify missing facts, evidence, urgency, and research steps.",
+                "",
+                640,
+            ],
+        ],
+    )
+    gr.Markdown(
+        f"Model: [{MODEL_ID}](https://huggingface.co/{MODEL_ID}) · "
+        f"Dataset: [{DATASET_ID}](https://huggingface.co/datasets/{DATASET_ID}) · "
+        "Adaption pairwise score: **92.16%** · Domain score: **65.66%**"
+    )
+
+demo.queue(default_concurrency_limit=1).launch()

requirements.txt

@@ -0,0 +1,9 @@
+accelerate>=1.7.0
+bitsandbytes>=0.45.5
+gradio==5.49.1
+huggingface_hub>=0.34.0
+peft>=0.15.1
+safetensors>=0.5.3
+torch>=2.6.0
+transformers>=4.51.3
+