#!/usr/bin/env python3 import csv import html import json import math import shutil import subprocess from collections import Counter, defaultdict from datetime import datetime from pathlib import Path ROOT = Path(__file__).resolve().parents[1] OUT_ROOT = ROOT / "presentation" PACKAGE = OUT_ROOT / "mwave_design_review_package" ASSETS = PACKAGE / "assets" CHARTS = ASSETS / "charts" DATA_DIR = PACKAGE / "source_data" METRICS_DIR = DATA_DIR / "metrics" PRED_DIR = DATA_DIR / "predictions" DERIVED_DIR = DATA_DIR / "derived" BOEING = "#0039A6" BOEING_2 = "#0067B1" RED = "#C8102E" BLACK = "#080A0F" INK = "#111827" STEEL = "#5B6472" MIST = "#E7ECF3" LIGHT = "#F7F9FC" COLORS = [BOEING, BOEING_2, RED, BLACK, STEEL, "#8EA4C8", "#B7C3D5", "#7F1D1D", "#334155", "#64748B"] METRIC_FILES = { "base_struct": ROOT / "outputs/metrics/base_struct_metrics.json", "finetuned_struct": ROOT / "outputs/metrics/finetuned_struct_metrics.json", "base_qa": ROOT / "outputs/metrics/base_qa_metrics.json", "finetuned_qa": ROOT / "outputs/metrics/finetuned_qa_metrics.json", } PRED_FILES = { "base_struct": ROOT / "outputs/predictions/base_struct_predictions.jsonl", "finetuned_struct": ROOT / "outputs/predictions/finetuned_struct_predictions.jsonl", "base_qa": ROOT / "outputs/predictions/base_qa_predictions.jsonl", "finetuned_qa": ROOT / "outputs/predictions/finetuned_qa_predictions.jsonl", } VAL_FILES = { "val_struct": ROOT / "data/processed/val_struct.jsonl", "val_qa": ROOT / "data/processed/val_qa.jsonl", "summary": ROOT / "data/processed/summary.json", } def read_json(path): return json.loads(Path(path).read_text(encoding="utf-8")) def read_jsonl(path): rows = [] with Path(path).open(encoding="utf-8") as f: for line in f: line = line.strip() if line: rows.append(json.loads(line)) return rows def safe_pct(v): if v is None: return "N/A" return f"{v * 100:.1f}%" def safe_num(v, digits=3): if v is None: return "N/A" if isinstance(v, float): if abs(v) < 0.001 and v != 0: return f"{v:.2e}" return f"{v:.{digits}f}".rstrip("0").rstrip(".") if isinstance(v, int): return f"{v:,}" return str(v) def esc(text): return html.escape(str(text), quote=True) def ensure_dirs(): if PACKAGE.exists(): shutil.rmtree(PACKAGE) for path in [CHARTS, METRICS_DIR, PRED_DIR, DERIVED_DIR]: path.mkdir(parents=True, exist_ok=True) def copy_source_files(): for name, src in METRIC_FILES.items(): shutil.copy2(src, METRICS_DIR / f"{name}_metrics.json") for name, src in PRED_FILES.items(): shutil.copy2(src, PRED_DIR / f"{name}_predictions.jsonl") for name, src in VAL_FILES.items(): shutil.copy2(src, DATA_DIR / src.name) adapter = ROOT / "outputs/qwen35_9b_lora" model_meta = { "base_model": "Qwen/Qwen3.5-9B", "adapter_dir": "outputs/qwen35_9b_lora", "adapter_files": sorted(p.name for p in adapter.glob("*") if p.is_file()) if adapter.exists() else [], "training_method": "4-bit QLoRA supervised fine-tuning", "train_file": "data/processed/train_mixed.jsonl", "validation_files": ["data/processed/val_struct.jsonl", "data/processed/val_qa.jsonl"], } (DATA_DIR / "model_training_metadata.json").write_text(json.dumps(model_meta, ensure_ascii=False, indent=2), encoding="utf-8") def write_csv(path, rows, fieldnames): with Path(path).open("w", encoding="utf-8", newline="") as f: writer = csv.DictWriter(f, fieldnames=fieldnames) writer.writeheader() writer.writerows(rows) def flatten_metrics(metrics): rows = [] for key, payload in metrics.items(): run = payload["run_name"] task = payload["task_type"] for metric, value in payload["metrics"].items(): rows.append({"run": run, "task": task, "metric": metric, "value": value}) return rows def class_stats(pred_rows, field="current_behavior"): labels = sorted({r.get("target", {}).get(field) for r in pred_rows if r.get("target", {}).get(field) is not None}) out = {} for label in labels: tp = fp = fn = correct = support = 0 for rec in pred_rows: target = rec.get("target", {}).get(field) pred = rec.get("prediction") if isinstance(rec.get("prediction"), dict) else {} pred_label = pred.get(field) if target == label: support += 1 if pred_label == label: correct += 1 if target == label and pred_label == label: tp += 1 elif target != label and pred_label == label: fp += 1 elif target == label and pred_label != label: fn += 1 precision = tp / (tp + fp) if tp + fp else 0.0 recall = tp / (tp + fn) if tp + fn else 0.0 f1 = 2 * precision * recall / (precision + recall) if precision + recall else 0.0 acc = correct / support if support else 0.0 out[label] = {"support": support, "accuracy": acc, "precision": precision, "recall": recall, "f1": f1} return out def confusion_rows(pred_rows, field="current_behavior", top_n=18): support = Counter(r.get("target", {}).get(field) for r in pred_rows) labels = [x for x, _ in support.most_common(top_n) if x is not None] matrix = [] for true_label in labels: row = {"true_label": true_label} total = 0 counts = Counter() for rec in pred_rows: pred = rec.get("prediction") if isinstance(rec.get("prediction"), dict) else {} if rec.get("target", {}).get(field) == true_label: total += 1 counts[pred.get(field)] += 1 for pred_label in labels: row[pred_label] = counts[pred_label] row["other"] = max(0, total - sum(row[p] for p in labels)) row["support"] = total matrix.append(row) return labels, matrix def qa_error_samples(rows, limit=20): samples = [] for idx, rec in enumerate(rows): target = rec.get("target", {}) pred = rec.get("prediction") if isinstance(rec.get("prediction"), dict) else {} errors = [] for key in ["occupied", "is_abnormal", "time_to_free_minutes", "used_areas"]: if target.get(key) != pred.get(key): errors.append(key) if errors: samples.append({ "index": idx, "error_fields": "|".join(errors), "target": json.dumps(target, ensure_ascii=False), "prediction": json.dumps(pred, ensure_ascii=False), }) if len(samples) >= limit: break return samples def derive_data(metrics, predictions): metric_rows = flatten_metrics(metrics) write_csv(DERIVED_DIR / "metric_comparison.csv", metric_rows, ["run", "task", "metric", "value"]) base_stats = class_stats(predictions["base_struct"]) ft_stats = class_stats(predictions["finetuned_struct"]) labels = sorted(set(base_stats) | set(ft_stats), key=lambda x: ft_stats.get(x, {}).get("support", 0), reverse=True) per_class = [] for label in labels: per_class.append({ "label": label, "support": ft_stats.get(label, {}).get("support", 0), "base_accuracy": base_stats.get(label, {}).get("accuracy", 0), "finetuned_accuracy": ft_stats.get(label, {}).get("accuracy", 0), "base_f1": base_stats.get(label, {}).get("f1", 0), "finetuned_f1": ft_stats.get(label, {}).get("f1", 0), "f1_delta": ft_stats.get(label, {}).get("f1", 0) - base_stats.get(label, {}).get("f1", 0), }) write_csv(DERIVED_DIR / "per_behavior_scores.csv", per_class, ["label", "support", "base_accuracy", "finetuned_accuracy", "base_f1", "finetuned_f1", "f1_delta"]) labels, cm = confusion_rows(predictions["finetuned_struct"]) write_csv(DERIVED_DIR / "finetuned_current_behavior_confusion_top18.csv", cm, ["true_label", *labels, "other", "support"]) qa_samples = qa_error_samples(predictions["finetuned_qa"]) write_csv(DERIVED_DIR / "finetuned_qa_error_samples.csv", qa_samples, ["index", "error_fields", "target", "prediction"]) metadata = { "generated_at": datetime.now().strftime("%Y-%m-%d %H:%M:%S"), "package_purpose": "Self-contained design review presentation for MWave aircraft lavatory radar LLM fine-tuning experiment.", "primary_audience": "Design experts and system reviewers", "storyline": [ "Problem: radar-derived lavatory behavior state is sequential, ambiguous, and privacy-sensitive.", "Intervention: fine-tune Qwen3.5-9B with mixed structured prediction and QA supervision.", "Evidence: compare base vs fine-tuned on identical validation sets.", "Design implication: edge-first state output should be robustly validated before aircraft integration.", ], "source_files": { "metrics": [str((METRICS_DIR / f"{name}_metrics.json").relative_to(PACKAGE)) for name in METRIC_FILES], "predictions": [str((PRED_DIR / f"{name}_predictions.jsonl").relative_to(PACKAGE)) for name in PRED_FILES], "derived": [str(p.relative_to(PACKAGE)) for p in sorted(DERIVED_DIR.glob("*"))], }, } (DATA_DIR / "presentation_metadata.json").write_text(json.dumps(metadata, ensure_ascii=False, indent=2), encoding="utf-8") return {"metric_rows": metric_rows, "per_class": per_class, "confusion_labels": labels, "confusion": cm, "metadata": metadata} def svg_wrap(width, height, title, desc, metadata, body): meta = esc(json.dumps(metadata, ensure_ascii=False, indent=2)) return f''' {esc(title)} {esc(desc)} {meta} {body} ''' def text(x, y, content, size=16, weight=500, fill=INK, anchor="start"): return f'{esc(content)}' def save_svg(name, width, height, title, desc, metadata, body): path = CHARTS / name path.write_text(svg_wrap(width, height, title, desc, metadata, body), encoding="utf-8") return path def bar_chart(name, title, subtitle, rows, value_key, label_key="label", group_key=None, width=980, height=560, value_format="percent"): margin = {"l": 110, "r": 46, "t": 106, "b": 110} plot_w = width - margin["l"] - margin["r"] plot_h = height - margin["t"] - margin["b"] max_v = max([r[value_key] for r in rows] + [1.0 if value_format == "percent" else 0.1]) if value_format == "percent": max_v = max(1.0, max_v) body = [ text(34, 42, title, 28, 900, BLACK), text(34, 72, subtitle, 14, 500, STEEL), f'', ] for i in range(5): y = margin["t"] + plot_h * i / 4 val = max_v * (1 - i / 4) body.append(f'') tick = safe_pct(val) if value_format == "percent" else safe_num(val, 1) body.append(text(margin["l"] - 14, y + 5, tick, 12, 600, STEEL, "end")) n = len(rows) slot = plot_w / max(n, 1) bar_w = slot * 0.62 for i, row in enumerate(rows): v = row[value_key] h = 0 if max_v == 0 else plot_h * v / max_v x = margin["l"] + i * slot + (slot - bar_w) / 2 y = margin["t"] + plot_h - h color = row.get("color") or COLORS[i % len(COLORS)] label = str(row[label_key]) meta = esc(json.dumps(row, ensure_ascii=False)) body.append(f'{esc(label)}: {safe_pct(v) if value_format == "percent" else safe_num(v)}') body.append(text(x + bar_w / 2, margin["t"] + plot_h + 28, label[:18], 11, 700, STEEL, "middle") + f'') if group_key: groups = [] last = None start = 0 for i, row in enumerate(rows + [{group_key: None}]): g = row.get(group_key) if last is None: last = g if g != last: groups.append((last, start, i - 1)) start = i last = g for group, start, end in groups: if group is None: continue x = margin["l"] + (start + end + 1) * slot / 2 body.append(text(x, height - 26, group, 13, 900, BLACK, "middle")) return save_svg(name, width, height, title, subtitle, {"chart_type": "bar", "rows": rows, "value_key": value_key}, "\n".join(body)) def grouped_metric_chart(metrics): order = [ ("Schema完整率", "required_field_complete_rate", "struct"), ("当前行为Acc", "current_behavior_accuracy", "struct"), ("当前行为Macro-F1", "current_behavior_macro_f1", "struct"), ("下一行为Acc", "next_possible_behavior_accuracy", "struct"), ("阶段Index Acc", "stage_index_accuracy", "struct"), ("序列Exact", "sequence_exact_match", "struct"), ("占用Acc", "occupied_accuracy", "qa"), ("异常F1", "is_abnormal_f1", "qa"), ("区域F1", "used_areas_micro_f1", "qa"), ] rows = [] for label, key, task in order: base = metrics[f"base_{task}"]["metrics"].get(key) ft = metrics[f"finetuned_{task}"]["metrics"].get(key) if base is not None: rows.append({"label": f"{label}\nBase", "metric": key, "run": "base", "task": task, "value": base, "color": STEEL}) if ft is not None: rows.append({"label": f"{label}\nFT", "metric": key, "run": "finetuned", "task": task, "value": ft, "color": BOEING if task == "struct" else RED}) return bar_chart( "01_base_vs_finetuned_scores.svg", "Base vs Fine-tuned: 指标总览", "同一验证集上的结构化预测与 QA 任务对比；柱形为可检查 metadata 的 SVG 元素。", rows, "value", width=1280, height=620, ) def label_distribution_chart(summary): rows = [] for label, count in sorted(summary["train_struct"]["label_counts"].items(), key=lambda x: x[1], reverse=True): rows.append({"label": label, "value": count, "color": BOEING if label == "坐用马桶" else COLORS[len(rows) % len(COLORS)]}) return bar_chart( "02_train_label_distribution.svg", "训练集行为标签分布", "类别分布显示任务主要由长时间状态和短过渡行为共同构成；设计评审应关注少数类边界。", rows[:19], "value", width=1280, height=620, value_format="count", ) def per_behavior_delta_chart(per_class): rows = [] for row in sorted(per_class, key=lambda x: x["f1_delta"], reverse=True)[:16]: rows.append({ "label": row["label"], "value": row["f1_delta"], "support": row["support"], "base_f1": row["base_f1"], "finetuned_f1": row["finetuned_f1"], "color": BOEING if row["f1_delta"] >= 0 else RED, }) return bar_chart( "03_behavior_f1_delta.svg", "行为类别 F1 改善幅度", "展示微调对各行为类别的收益；support 表示验证集该行为样本量。", rows, "value", width=1280, height=620, ) def time_mae_chart(metrics): rows = [] mapping = [ ("当前行为已持续", "elapsed_seconds_in_current_behavior_mae"), ("当前行为剩余", "estimated_remaining_seconds_mae"), ("完整流程剩余", "full_remaining_seconds_mae"), ("当前行为结束时刻", "expected_end_time_mae"), ] for label, key in mapping: v = metrics["finetuned_struct"]["metrics"].get(key) if v is not None: rows.append({"label": label, "metric": key, "value": v, "color": BOEING}) rows.append({"label": "QA空出时间(分钟)", "metric": "time_to_free_minutes_mae", "value": metrics["finetuned_qa"]["metrics"]["time_to_free_minutes_mae"], "color": RED}) return bar_chart( "04_time_error_mae.svg", "时间估计误差", "结构化任务以秒为单位；QA 空出时间以分钟为单位，接近 0 说明该 QA 目标在当前构造规则下被模型很好拟合。", rows, "value", width=1100, height=560, value_format="count", ) def confusion_heatmap(name, title, labels, matrix): width, height = 1180, 900 margin = {"l": 170, "r": 44, "t": 126, "b": 170} n = len(labels) cell = min((width - margin["l"] - margin["r"]) / n, (height - margin["t"] - margin["b"]) / n) max_count = max([row[label] for row in matrix for label in labels] + [1]) body = [ text(34, 44, title, 28, 900, BLACK), text(34, 74, "Fine-tuned current_behavior confusion matrix, top validation labels.", 14, 500, STEEL), f'', ] for i, true_label in enumerate(labels): body.append(text(margin["l"] - 12, margin["t"] + i * cell + cell * 0.62, true_label, 12, 700, INK, "end")) body.append(text(margin["l"] + i * cell + cell * 0.5, margin["t"] + n * cell + 26, true_label, 11, 700, INK, "middle")) for i, row in enumerate(matrix): for j, pred_label in enumerate(labels): count = row[pred_label] intensity = count / max_count blue = int(246 - 170 * intensity) color = f"rgb({blue},{max(39, blue + 12)},{255})" if i == j else f"rgb({255},{max(230, 248 - int(120 * intensity))},{max(230, 248 - int(120 * intensity))})" meta = esc(json.dumps({"true_label": row["true_label"], "predicted_label": pred_label, "count": count, "support": row["support"]}, ensure_ascii=False)) x = margin["l"] + j * cell y = margin["t"] + i * cell body.append(f'true={esc(row["true_label"])}, pred={esc(pred_label)}, count={count}') if count: body.append(text(x + cell / 2, y + cell * 0.62, count, 9, 700, BLACK, "middle")) body.append(text(margin["l"] + cell * n / 2, height - 38, "Predicted behavior", 14, 900, BLACK, "middle")) body.append(f'True behavior') return save_svg(name, width, height, title, "Confusion heatmap with metadata on each cell.", {"chart_type": "confusion_matrix", "labels": labels}, "\n".join(body)) def architecture_svg(): width, height = 1280, 500 boxes = [ ("毫米波雷达", "点云/轨迹窗口\n隐私友好，非视觉图像", 56), ("边缘特征层", "去噪、跟踪、区域、速度\n生成结构化中间表征", 356), ("Qwen3.5-9B LoRA", "本地推理输出 JSON\n行为、阶段、剩余时间、QA", 656), ("客舱系统", "乘务终端、维护日志\n异常关注而非强控制", 956), ] body = [text(34, 44, "飞机厕所部署技术方案", 30, 900, BLACK), text(34, 76, "Edge-first, privacy-preserving, fail-safe integration concept.", 14, 500, STEEL)] for i, (title, desc, x) in enumerate(boxes): body.append(f'') body.append(text(x + 24, 190, title, 22, 900, BOEING if i < 3 else RED)) for k, line in enumerate(desc.split("\n")): body.append(text(x + 24, 238 + k * 28, line, 16, 600, STEEL)) if i < len(boxes) - 1: body.append(f'') body.append(f'') body.append(text(56, 420, "设计原则：厕所侧只上传结构化状态，不上传可逆原始人体数据；LLM JSON 校验失败时回退规则模型。", 18, 800, BLACK)) return save_svg("06_deployment_architecture.svg", width, height, "飞机厕所部署技术方案", "Edge deployment architecture.", {"chart_type": "architecture"}, "\n".join(body)) def build_svgs(summary, metrics, derived): paths = [ grouped_metric_chart(metrics), label_distribution_chart(summary), per_behavior_delta_chart(derived["per_class"]), time_mae_chart(metrics), confusion_heatmap("05_finetuned_behavior_confusion.svg", "Fine-tuned 行为混淆矩阵", derived["confusion_labels"], derived["confusion"]), architecture_svg(), ] chart_manifest = [] for path in paths: chart_manifest.append({ "file": str(path.relative_to(PACKAGE)), "title": path.stem.replace("_", " "), "metadata_note": "Open the SVG file in a text editor or use the HTML metadata explorer to inspect embedded and per-element data-meta.", }) (CHARTS / "chart_manifest.json").write_text(json.dumps(chart_manifest, ensure_ascii=False, indent=2), encoding="utf-8") return chart_manifest def metric_delta(metrics, task, key): base = metrics[f"base_{task}"]["metrics"].get(key) ft = metrics[f"finetuned_{task}"]["metrics"].get(key) if base is None or ft is None: return None return ft - base def build_html(summary, metrics, derived, chart_manifest): struct = metrics["finetuned_struct"]["metrics"] qa = metrics["finetuned_qa"]["metrics"] base_struct = metrics["base_struct"]["metrics"] base_qa = metrics["base_qa"]["metrics"] metadata_json = json.dumps({ "summary": summary, "metrics": metrics, "chart_manifest": chart_manifest, "derived_files": [str(p.relative_to(PACKAGE)) for p in sorted(DERIVED_DIR.glob("*"))], }, ensure_ascii=False) metric_table_rows = [] story_metrics = [ ("结构化 schema 完整率", base_struct["required_field_complete_rate"], struct["required_field_complete_rate"], "决定输出能否进入机载系统状态总线。"), ("当前行为准确率", base_struct["current_behavior_accuracy"], struct["current_behavior_accuracy"], "核心状态识别能力，从 radar 表征映射到行为标签。"), ("当前行为 Macro-F1", base_struct["current_behavior_macro_f1"], struct["current_behavior_macro_f1"], "衡量少数类是否被覆盖，适合设计评审关注长尾行为。"), ("下一行为准确率", base_struct["next_possible_behavior_accuracy"], struct["next_possible_behavior_accuracy"], "影响短期流程预测和预计空出时间。"), ("阶段 index 准确率", base_struct["stage_index_accuracy"], struct["stage_index_accuracy"], "反映模型是否理解完整厕所使用流程位置。"), ("QA 异常 F1", base_qa["is_abnormal_f1"], qa["is_abnormal_f1"], "用于乘务关注提示，需单独评估。"), ("QA 区域 F1", base_qa["used_areas_micro_f1"], qa["used_areas_micro_f1"], "用于理解马桶、洗手池、门、垃圾桶等区域使用状态。"), ] for name, base, ft, meaning in story_metrics: delta = ft - base if base is not None and ft is not None else None metric_table_rows.append(f"{esc(name)}{safe_pct(base)}{safe_pct(ft)}{safe_pct(delta) if delta is not None else 'N/A'}{esc(meaning)}") top_improvements = sorted(derived["per_class"], key=lambda x: x["f1_delta"], reverse=True)[:8] top_rows = "\n".join( f"{esc(r['label'])}{r['support']}{safe_pct(r['base_f1'])}{safe_pct(r['finetuned_f1'])}{safe_pct(r['f1_delta'])}" for r in top_improvements ) chart_cards = "\n".join( f'''

{esc(item["title"])}

{esc(item["metadata_note"])}

''' for item in chart_manifest ) source_links = "\n".join( f"
  • {esc(str(p.relative_to(PACKAGE)))}
    • " for p in sorted(DATA_DIR.rglob("*")) if p.is_file() ) html_text = f''' MWave Design Review Presentation
    MWave Radar LLM · Aircraft Lavatory Design Review

    从毫米波雷达到机载厕所状态智能：一次可验证的 LLM 微调实验

    这份汇报面向设计专家评审，重点不是展示训练日志，而是讲清楚：为什么需要模型、模型学到了什么、哪些结果能支持设计决策、哪些风险需要在机载部署前继续验证。

    {summary["train_struct"]["num_examples"]:,}
    训练结构样本
    {summary["val_struct"]["num_examples"]:,}
    验证结构样本
    {summary["train_qa"]["num_examples"]:,}
    QA 训练样本
    Qwen3.5-9B
    4-bit QLoRA 微调

    评审 Story Line

    1

    设计问题

    飞机厕所内不能依赖摄像头，毫米波雷达提供隐私友好的时空轨迹，但原始序列难以直接变成可用的客舱状态。

    2

    模型角色

    LLM 不替代雷达信号处理，而是读取结构化窗口和中间表征，输出严格 JSON：当前行为、下一行为、阶段、剩余时间、序列，以及 QA 状态。

    3

    实验判断

    同一验证集上比较 base Qwen3.5-9B 与微调后模型，指标覆盖 schema 合规、行为识别、流程理解、时间估计和 QA 推理。

    4

    设计结论

    微调显著提升结构化输出和 QA 可用性，但时间误差、少数类行为和异常提示仍应进入下一轮系统设计验证。

    任务与评估方法

    结构化预测

    输入雷达时序窗口和中间层表征，输出 current_behavior、next_possible_behavior、stage_index、remaining time、sequence_so_far 等字段。

    QA 状态回答

    作为独立维度评估 occupied、time_to_free_minutes、used_areas、is_abnormal，避免只看结构化任务而忽略最终用户问题。

    指标选择

    分类用 accuracy/F1，schema 用 JSON parse 和 required field complete，时间用 MAE，序列用 exact/prefix/last-label match。

    关键结果

    {safe_pct(struct["current_behavior_accuracy"])}
    当前行为准确率

    Base 为 {safe_pct(base_struct["current_behavior_accuracy"])}。

    {safe_pct(struct["required_field_complete_rate"])}
    结构化字段完整率

    Base 为 {safe_pct(base_struct["required_field_complete_rate"])}，说明微调主要解决 schema 对齐。

    {safe_pct(qa["is_abnormal_f1"])}
    QA 异常 F1

    Base 为 {safe_pct(base_qa["is_abnormal_f1"])}。

    {safe_num(struct["full_remaining_seconds_mae"], 1)}s
    完整流程剩余时间 MAE

    该误差直接影响预计空出时间体验。

    {''.join(metric_table_rows)}
    评估项BaseFine-tuned变化设计含义
    评审判断：当前实验已经证明 9B LoRA 可以把通用 LLM 拉到可用的结构化状态输出轨道上；但机载上线前仍应增加跨乘客、跨机型、跨雷达安装角度的数据验证，并对异常提示设定保守阈值。

    哪些行为受益最大

    下表展示 F1 改善最大的行为类别。设计评审时应同时看 support，避免把少量样本上的提升误判为稳定能力。

    {top_rows}
    行为验证样本数Base F1Fine-tuned F1提升

    SVG 矢量图表

    所有图表都在 assets/charts/ 下以 SVG 保存，可缩放、可放进设计文档；SVG 内包含 metadata，HTML 下方可查看。

    {chart_cards}

    Metadata Explorer

    选择图表或数据对象，查看生成依据、源文件路径和图表内嵌 metadata。此功能完全离线，不依赖外网。

    选择对象

    提示：SVG 图表也可以直接用文本编辑器打开，查看其中的 <metadata> 和每个 bar/cell 的 data-meta。

    
    
    
  
    

  
    
    
    部署到飞机厕所的技术方案
    
    
    
      
    边缘优先
    厕所侧完成雷达预处理和特征抽取，客舱边缘计算单元加载量化 LLM/LoRA，减少原始数据移动。
    
      
    状态输出
    对外只输出 occupied、estimated free time、used areas、abnormal flag 和行为阶段，不输出身份或可逆人体点云。
    
      
    失效保护
    JSON 校验失败、置信不足或异常连续触发时回退规则模型，并只给出“需关注”级提示。
    
    
    
    
  
    

  
    
    
    可拷走的源数据与材料
    
    
    整个 mwave_design_review_package 目录可直接复制到其他电脑。HTML、PDF、SVG 和 CSV/JSONL 都使用相对路径或独立文件。
    
    
      {source_links}
    
  
    

  
    Generated locally · {datetime.now().strftime("%Y-%m-%d %H:%M:%S")} · Boeing blue / red / black / gray design system
    

    




'''
    (PACKAGE / "index.html").write_text(html_text, encoding="utf-8")


def build_readme():
    readme = f"""# MWave Design Review Presentation Package

This folder is self-contained and can be copied to another computer.

Open:
- `index.html`: interactive presentation with SVG charts and metadata explorer.
- `mwave_design_review_report.pdf`: static PDF export for review circulation.

Key folders:
- `assets/charts/`: standalone SVG vector charts. Each SVG includes embedded metadata.
- `source_data/metrics/`: original evaluation metric JSON files.
- `source_data/predictions/`: original prediction JSONL files for base and fine-tuned runs.
- `source_data/derived/`: CSV files derived for the presentation, including per-behavior scores and confusion matrix.

Generated at: {datetime.now().strftime("%Y-%m-%d %H:%M:%S")}
"""
    (PACKAGE / "README.md").write_text(readme, encoding="utf-8")


def export_pdf():
    html_path = PACKAGE / "index.html"
    pdf_path = PACKAGE / "mwave_design_review_report.pdf"
    chrome = shutil.which("google-chrome") or shutil.which("chromium") or shutil.which("chromium-browser")
    if not chrome:
        print("Chrome not found; skipped PDF export.")
        return False
    cmd = [
        chrome,
        "--headless",
        "--no-sandbox",
        "--disable-gpu",
        "--print-to-pdf=" + str(pdf_path),
        "--print-to-pdf-no-header",
        "file://" + str(html_path.resolve()),
    ]
    subprocess.run(cmd, check=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
    return True


def make_zip():
    zip_base = OUT_ROOT / "mwave_design_review_package"
    archive = shutil.make_archive(str(zip_base), "zip", root_dir=OUT_ROOT, base_dir=PACKAGE.name)
    return Path(archive)


def main():
    ensure_dirs()
    copy_source_files()
    summary = read_json(ROOT / "data/processed/summary.json")
    metrics = {key: read_json(path) for key, path in METRIC_FILES.items()}
    predictions = {key: read_jsonl(path) for key, path in PRED_FILES.items()}
    derived = derive_data(metrics, predictions)
    chart_manifest = build_svgs(summary, metrics, derived)
    build_html(summary, metrics, derived, chart_manifest)
    build_readme()
    pdf_ok = export_pdf()
    archive = make_zip()
    print(f"wrote package: {PACKAGE}")
    print(f"wrote archive: {archive}")
    print(f"pdf_exported: {pdf_ok}")


if __name__ == "__main__":
    main()