build-small-hackathon/InContext

1import gradio as gr2import torch3import json4import html5import traceback6from transformers import AutoModelForCausalLM, AutoTokenizer7 8print("Loading model...")9model_name = "Qwen/Qwen2.5-0.5B-Instruct"10tokenizer = AutoTokenizer.from_pretrained(model_name)11model = AutoModelForCausalLM.from_pretrained(12    model_name,13    torch_dtype=torch.float16,14    device_map="auto"15)16print("Model loaded.")17 18SYSTEM_PROMPT = """You are an English learning assistant. Extract 8-20 useful expressions from the text.19For each expression, output a JSON object with keys: expression, meaning, explanation, original_context, extra_example.20Meaning and explanation should be in Chinese.21Output must be a JSON array. No extra text."""22 23def analyze(text):24    try:25        if not text or len(text.strip()) < 20:26            return "<div style='color:red'>⚠️ Please enter at least 20 characters.</div>"27 28        messages = [29            {"role": "system", "content": SYSTEM_PROMPT},30            {"role": "user", "content": text}31        ]32        inputs = tokenizer.apply_chat_template(33            messages,34            add_generation_prompt=True,35            return_tensors="pt"36        ).to(model.device)37 38        with torch.no_grad():39            outputs = model.generate(40                inputs,41                max_new_tokens=1024,42                do_sample=False,43                temperature=1.044            )45 46        response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)47 48        # 提取 JSON49        if "```json" in response:50            response = response.split("```json")[1].split("```")[0]51        elif "```" in response:52            response = response.split("```")[1].split("```")[0]53        start = response.find("[")54        end = response.rfind("]") + 155        if start == -1 or end == 0:56            return f"<div style='color:red'>No JSON array found. Raw response:<br>{html.escape(response[:300])}</div>"57 58        json_str = response[start:end]59        data = json.loads(json_str)60 61        cards = ""62        for e in data:63            cards += f"""64            <div style="background:white;border-radius:16px;border:1px solid #ddd;padding:1rem;margin-bottom:1rem;">65                <b style="font-size:1.2rem;">{html.escape(str(e.get('expression', '')))}</b><br>66                <b>Meaning</b><br>{html.escape(str(e.get('meaning', '')))}<br>67                <b>Explanation</b><br>{html.escape(str(e.get('explanation', '')))}<br>68                <b>Original Context</b><br>{html.escape(str(e.get('original_context', '')))}<br>69                <b>Extra Example</b><br>{html.escape(str(e.get('extra_example', '')))}70            </div>71            """72        return cards if cards else "<div>No expressions extracted.</div>"73    except Exception as e:74        error_html = f"<div style='color:red; background:#ffe0e0; padding:1rem; border-radius:8px;'>"75        error_html += f"<b>Error:</b> {html.escape(str(e))}<br><br>"76        error_html += f"<details><summary>Full traceback</summary><pre>{html.escape(traceback.format_exc())}</pre></details>"77        error_html += "</div>"78        return error_html79 80# 浅色主题81theme = gr.themes.Soft(82    primary_hue="neutral",83    secondary_hue="neutral",84    font=gr.themes.GoogleFont("Inter"),85).set(86    body_background_fill="#fafaf9",87    button_primary_background_fill="#1a1a1a",88    button_primary_text_color="white",89    block_background_fill="white",90)91 92with gr.Blocks(theme=theme, title="InContext") as demo:93    gr.Markdown("# InContext\n### Learn English Expressions Through Real Content")94    with gr.Row():95        txt = gr.Textbox(lines=10, placeholder="Paste English content here...", label="")96    btn = gr.Button("Analyze", variant="primary")97    out = gr.HTML()98    btn.click(analyze, txt, out)99 100demo.launch()