diff --git a/evals/check/build.py b/evals/check/build.py index c86bd4b..13a127a 100644 --- a/evals/check/build.py +++ b/evals/check/build.py @@ -58,6 +58,7 @@ BIDDER = { "agent": "陈立新", "date": "2026年11月20日", } +NUMS = re.compile(r"\d+(?:\.\d+)?") DIGITS = "零壹贰叁肆伍陆柒捌玖" UNITS = ["", "拾", "佰", "仟"] SECTIONS = ["", "万", "亿"] @@ -181,6 +182,7 @@ def _row(k: int, r: dict[str, Any], columns: list[str]) -> dict[str, Any]: "responds": r["responds"], "req": r["req"], "iid": r["iid"], + "key": r.get("key"), "tp": r["tp"], } @@ -219,10 +221,13 @@ def build(tender: str, bid: str, style: str, run: str) -> None: answer, dev = comply(raw, positive) else: # 功能类要求：复述型（风格 A）照常写「满足，……」 answer, dev = "满足，" + re.sub(r"^[\s★▲*＊\d.、（）()]+", "", req_text)[:120], "无偏离" - rw = rewrite.get(iid) or {} + rw = dict(rewrite.get(iid) or {}) + if "answer" in rw and NUMS.findall(rw["answer"]) != NUMS.findall(answer): + rw.pop("answer") # 改写后的数字和规矩的响应对不上：丢弃改写，规矩靠构造保证 rows_by_route.setdefault(route, []).append( { "iid": iid, + "key": iid, "requirement": rw.get("requirement", req_text), "answer": rw.get("answer", answer), "deviation": dev, @@ -253,11 +258,13 @@ def build(tender: str, bid: str, style: str, run: str) -> None: if not sents: continue text = "".join(x.text for x in sents) + rw = rewrite.get(f"u:{u.id}") or {} rows_by_route.setdefault("tech_row", []).append( { "iid": None, - "requirement": text, - "answer": "满足：" + text, + "key": f"u:{u.id}", + "requirement": rw.get("requirement", text), + "answer": rw.get("answer", "满足：" + text), "deviation": "无偏离", "responds": [u.id], "req": {"p": sents[0].page_no, "t": gold.norm(sents[0].text)[:24]}, @@ -298,7 +305,10 @@ def build(tender: str, bid: str, style: str, run: str) -> None: lines = [(total * b / sum(base)).quantize(Decimal("0.01")) for b in base] lines[-1] = total - sum(lines[:-1]) unit_prices = [(ln / w).quantize(Decimal("0.01")) for ln, w in zip(lines, weights, strict=False)] + moved: list[dict[str, Any]] = [] # 风格 B：一部分技术行写在商务资格册里（干扰项 T01） + moved_columns: list[str] = [] for n, f in enumerate(profile["forms"], start=1): + start = len(blocks) head = {"type": "heading", "text": f"{n}．{f['title']}", "new_page": n > 1, "form": f["key"]} kind = f["kind"] if kind == "text": @@ -364,6 +374,15 @@ def build(tender: str, bid: str, style: str, run: str) -> None: else: rows = [_row(k, r, columns) for k, r in enumerate(src, 1)] pool = [] + if style == "B" and f.get("route", "tech_row") == "tech_row" and not f.get("mode"): + keep = [] + for k, r in enumerate(rows): + (moved if k % 12 == 11 else keep).append(r) + rows, moved_columns = keep, columns + if style == "B" and "no" in columns: # 风格 B 用自己的编号 + prefix = "技" if f.get("route", "tech_row") == "tech_row" else "商" + for k, r in enumerate(rows, 1): + r["cells"][columns.index("no")] = f"{prefix}-{k:02d}" blocks += [ head, { @@ -373,6 +392,7 @@ def build(tender: str, bid: str, style: str, run: str) -> None: "form": f["key"], "route": f.get("route", "tech_row"), "only_deviations": bool(f.get("only_deviations")), + "mode": f.get("mode"), "columns": columns, "pool": pool, }, @@ -399,6 +419,28 @@ def build(tender: str, bid: str, style: str, run: str) -> None: blocks.append(b) else: raise ValueError(f"未知的表单类型：{kind}") + tech = kind == "response_table" and f.get("route", "tech_row") == "tech_row" + for b in blocks[start:]: + b["vol"] = "v2" if style == "B" and tech else "v1" + if moved: + for k, r in enumerate(moved, 1): + if "no" in moved_columns: + r["cells"][moved_columns.index("no")] = f"补-{k:02d}" + blocks += [ + {"type": "heading", "text": "补充技术响应", "new_page": True, "vol": "v1"}, + { + "type": "table", + "header": moved_columns, + "rows": moved, + "form": "moved", + "route": "tech_row", + "only_deviations": False, + "mode": None, + "columns": moved_columns, + "pool": [], + "vol": "v1", + }, + ] for sec in profile.get("prose", []): blocks.append({"type": "heading", "text": sec["title"], "new_page": True}) units = sorted(set(form_responds.get("prose:" + sec["title"], []) + form_responds.get("prose:", []))) @@ -406,6 +448,29 @@ def build(tender: str, bid: str, style: str, run: str) -> None: {"type": "para", "text": fill(p, ctx), "responds": units if i == 0 else [], "form": "prose"} for i, p in enumerate(sec["paragraphs"]) ] + for b in blocks[-len(sec["paragraphs"]) - 1 :]: + b["vol"] = "v2" if style == "B" else "v1" + blocks[0]["vol"] = "v1" + if style == "B": + volumes = [ + {"id": "v1", "kind": "commercial", "header": None, "blocks": [b for b in blocks if b.get("vol") == "v1"]}, + { + "id": "v2", + "kind": "technical", + "header": None, + "blocks": [ + {"type": "title", "text": f"{proj.get('name', '')}投标文件（技术册）"}, + *[b for b in blocks if b.get("vol") == "v2"], + ], + }, + ] + else: + volumes = [{"id": "v1", "kind": "whole", "header": None, "blocks": blocks}] + traps = {u: "T01" for r in moved for u in r.get("responds") or []} + for v in rewrite: # 被改写的要求（截断、同义替换）是 T02；只写值不复述是 T03 + for r in [x for xs in rows_by_route.values() for x in xs if x.get("key") == v]: + for u in r["responds"]: + traps.setdefault(u, "T03" if "answer" in rewrite[v] and "requirement" not in rewrite[v] else "T02") spec = { "bid": bid, @@ -413,7 +478,8 @@ def build(tender: str, bid: str, style: str, run: str) -> None: "style": style, "project_name": proj.get("name", ""), "bidder": BIDDER, - "volumes": [{"id": "v1", "kind": "whole", "header": None, "blocks": blocks}], + "volumes": volumes, + "traps": traps, "price": { "total": f"{total:.2f}", "items": [ diff --git a/evals/check/mutate.py b/evals/check/mutate.py index 31ddf0e..a1fb8f7 100644 --- a/evals/check/mutate.py +++ b/evals/check/mutate.py @@ -29,6 +29,7 @@ from build import rmb_upper # noqa: E402 OUT = EVALS / "out" / "check" VOCAB = {"dev": EVALS / "check" / "vocab-dev.json", "holdout": OUT / "_sealed" / "vocab-holdout.json"} +DOC_REQ = re.compile(r"提供.{0,12}(方案|证明|报告|文件|承诺|截图|复印件|证书)") UNIT_DOWN = {"TB": ("GB", 1024), "GB": ("MB", 1024), "Gbps": ("Mbps", 1000), "s": ("ms", 1000), "秒": ("毫秒", 1000)} @@ -37,7 +38,23 @@ class Mutator: self.spec, self.vocab, self.rng = spec, vocab, rng self.unit_defects: dict[str, dict[str, Any]] = {} self.rule_defects: list[dict[str, Any]] = [] - self.blocks = spec["volumes"][0]["blocks"] + self.vols = spec["volumes"] + + @property + def blocks(self) -> list[dict[str, Any]]: + """所有分册的块（只读视图）；增删用 remove / insert_after。""" + return [b for v in self.vols for b in v["blocks"]] + + def remove(self, b: dict[str, Any]) -> None: + for v in self.vols: + if b in v["blocks"]: + v["blocks"].remove(b) + + def insert_after(self, anchor: dict[str, Any], new: dict[str, Any]) -> None: + for v in self.vols: + if anchor in v["blocks"]: + v["blocks"].insert(v["blocks"].index(anchor) + 1, new) + return # ---------- 定位 ---------- def tables(self, *, explicit: bool | None = None, fixed: bool = False) -> list[dict[str, Any]]: @@ -51,7 +68,8 @@ class Mutator: continue if b.get("fixed") or "only_deviations" not in b: continue - if explicit is None or b["only_deviations"] != explicit: + mode = table_mode(b) + if explicit is None or (explicit and mode == "explicit") or (explicit is False and mode == "implicit"): out.append(b) return out @@ -118,6 +136,12 @@ class Mutator: return False tp = r.get("tp") or {} units = r.get("responds") or [] + req_text = r["cells"][req] if req is not None else "" + if DOC_REQ.search(req_text): # 「提供……方案 / 证明 / 报告」是文件要在（presence），不注入参数类问题 + return False + numeric = tp.get("op") in (">=", "<=", "=") and tp.get("value") + if op == "D05" and numeric: # 「暂不支持」只套在功能类要求上（尺寸、费用这类数值参数不会这么写） + return False if op == "D01": t["rows"].remove(r) self.drop_everywhere(units) @@ -180,7 +204,13 @@ class Mutator: def implicit_deviation(self, n: int) -> None: """只列偏离的表：挑默认接受的要求写成偏离（D06 负偏离未说明 / D05 写明不支持）。""" for t in self.tables(explicit=False): - pool = [r for r in t.get("pool") or [] if self.free(r.get("responds") or [])] + req_col = self.col(t, "requirement") + pool = [ + r + for r in t.get("pool") or [] + if self.free(r.get("responds") or []) + and not (req_col is not None and DOC_REQ.search(r["cells"][req_col])) + ] self.rng.shuffle(pool) ans, dev, note, req = (self.col(t, f) for f in ("answer", "deviation", "note", "requirement")) if ans is None or dev is None: @@ -192,10 +222,15 @@ class Mutator: if note is not None: row["cells"][note] = "" self.mark(row["responds"], "D06", row["cells"][dev]) + elif (tp := row.get("tp") or {}).get("op") in (">=", "<=") and tp.get("value"): + # 数值参数不会写「暂不支持」：写一个不满足的值并标负偏离 + v = Decimal(str(tp["value"])) + bad = v / 2 if tp["op"] == ">=" else v * Decimal("1.6") + row["cells"][ans] = f"{tp.get('param') or ''}：{bad.normalize():f}{tp.get('unit') or ''}" + row["cells"][dev] = "负偏离" + self.mark(row["responds"], "D06", "数值不满足并写负偏离") else: - param = (row.get("tp") or {}).get("param") or ( - row["cells"][req][:12] if req is not None else "该功能" - ) + param = tp.get("param") or (row["cells"][req][:12] if req is not None else "该功能") row["cells"][ans] = self.phrase("D05", "phrases").format(param=param) row["cells"][dev] = "负偏离" self.mark(row["responds"], "D05", row["cells"][ans]) @@ -214,7 +249,7 @@ class Mutator: kinds = {cats.get(u) for u in units} defect = "D11" if "qualification" in kinds else "D13" if "scoring" in kinds else "D02" self.mark(units, defect, "附件缺失", worded=False) - self.blocks.remove(b) + self.remove(b) def compliance(self) -> None: for t in self.tables(fixed=True): @@ -259,7 +294,7 @@ class Mutator: def p06(self) -> None: fp = self.rng.choice(self.vocab["P06"]["foreign_projects"]) - self.spec["volumes"][0]["header"] = fp["name"] + self.vols[-1]["header"] = fp["name"] # 技术分册照搬旧标书最常见 self.rule("P06", "页眉留旧项目名", fp["name"]) def text_paras(self, form_hint: str) -> list[dict[str, Any]]: @@ -295,7 +330,7 @@ class Mutator: letter = next((i for i, b in enumerate(self.blocks) if b.get("type") == "para" and b.get("form")), None) if letter is not None: text = self.phrase("P10", "phrases") - self.blocks.insert(letter + 1, {"type": "para", "text": text, "form": self.blocks[letter]["form"]}) + self.insert_after(self.blocks[letter], {"type": "para", "text": text, "form": self.blocks[letter]["form"]}) self.rule("P10", text, text) def p12(self) -> None: @@ -303,18 +338,27 @@ class Mutator: if t: alt = f"{Decimal(self.spec['price']['total']) * Decimal('0.97'):,.2f}" text = self.phrase("P12", "phrases").format(alt=alt) - self.blocks.insert(self.blocks.index(t) + 1, {"type": "para", "text": text}) + self.insert_after(t, {"type": "para", "text": text}) self.rule("P12", text, text) +def table_mode(b: dict[str, Any]) -> str: + """表格的响应方式：档案里表单写了 mode 的以它为准（如 S02 评分要的技术规格表是 presence，不是逐条响应的机制）。""" + return str(b.get("mode") or ("implicit" if b.get("only_deviations") else "explicit")) + + def unit_modes(spec: dict[str, Any]) -> dict[str, str]: + """单元的响应方式：逐条响应表里有 → explicit；只在只列偏离的表里 → implicit；附件 → evidence；格式文件 → presence。""" modes: dict[str, str] = {} - for b in spec["volumes"][0]["blocks"]: + rank = {"explicit": 0, "implicit": 1, "evidence": 2, "presence": 3} + blocks = [b for v in spec["volumes"] for b in v["blocks"]] + for b in blocks: if b.get("type") == "table" and "only_deviations" in b: - mode = "implicit" if b["only_deviations"] else "explicit" + mode = table_mode(b) for r in [*b["rows"], *(b.get("pool") or [])]: for u in r.get("responds") or []: - modes.setdefault(u, mode) + if u not in modes or rank.get(mode, 9) < rank.get(modes[u], 9): + modes[u] = mode elif b.get("attach"): for u in b.get("responds") or []: modes.setdefault(u, "evidence") @@ -366,6 +410,8 @@ def main() -> None: entry.update({"expect": "defect", "reviewed": False, **m.unit_defects[u.id]}) elif mode == "implicit": entry.update({"expect": "ok", "trap": "T06"}) + elif u.id in (spec.get("traps") or {}): + entry.update({"expect": "ok", "trap": spec["traps"][u.id]}) else: entry["expect"] = {"na": "na", "unknown": "unknown"}.get(mode, "ok") manifest_units[u.id] = entry diff --git a/evals/check/prompts/profile.md b/evals/check/prompts/profile.md new file mode 100644 index 0000000..343c2ca --- /dev/null +++ b/evals/check/prompts/profile.md @@ -0,0 +1,44 @@ +# 写「招标文件档案」的说明 + +为查标评测集写一份招标文件档案（tender profile），供 `evals/check/build.py` 生成虚构的规矩投标文件。 + +## 读这些 + +- 档案格式：`evals/check/build.py` 文件头的 docstring，以及各 kind 的生成逻辑。 +- 响应方式的定义：`evals/check/GUIDELINES.md` §2（含本文件那一行的记录，记录不准时以招标原文为准，并在回复里说明）。 +- 招标文件全文（带句子编号 `[p页s句]`）：`evals/out/docs//sentences.txt` +- 读标正式清单（要分配 routes 的条目）：`evals/out/check/tenders/.items.tsv`（列：iid、类别、重要性、页码、标题） +- 写好的档案可以参考格式（不要照抄内容）：`evals/out/check/tenders/S02.json`、`S04.json`、`S10.json`。 + +**不要读** `evals/out/check/_sealed/`、`server/`、`prompts/`（仓库根目录那个）。 + +## 只写一个文件 + +`evals/out/check/tenders/.json`。不改仓库里的其他文件，不做 git 操作。 + +## 要求 + +1. `project`：项目名称、编号、采购人、代理机构、投标截止日期（`deadline`，YYYY-MM-DD）、投标有效期天数、最高限价或预算（元，纯数字字符串）、采购包（编号、名称、限价）。多包时只投一个包，写 `bidding_lot`，`ceiling_total` 填那个包的限价，只属于别的包的条目 routes 记 `na`。全部从招标原文取。 +2. `modes`：按类别（tech_param、substantive、commercial、qualification、scoring，必要时加 scope）记响应方式（explicit / implicit / evidence / presence / platform），evidence 写 `when`（bid 或 award）。每条带 `basis`：`p` 页码，`t` 从 sentences.txt 逐字复制、不少于 8 个字的片段（去掉 `` 这类标记）。同一份文件里两处写法冲突时都记上，并在回复里说明。 +3. `forms`：按「投标文件格式」章的顺序列出投标人要交的全部格式文件。 + - 叙述性的（投标函、授权委托书、声明函、承诺书）用 `kind: "text"`，`paragraphs` 照抄固定文字，填空处换成占位符（见 docstring），没有对应占位符的填空写成虚构的合理内容。声明函按规矩投标人的正确填法写；「如有 / 如是」而规矩投标人不适用的不列。 + - 开标一览表 `price_summary`；分项报价表 `price_items`，`items` 写名称、数量、单位。**有分项最高限价或单价上限的，每项加 `ceiling`（元）**；上限是单价时再加 `ceiling_is_unit: true`。 + - 偏离表 / 响应表 `response_table`：`route` 是 `tech_row` 或 `commercial_row`；implicit 的写 `only_deviations: true`，explicit 的写 false。表头照抄格式（脚本按表头文字决定每一列放什么，支持「序号 / 名称 / 要求 / 响应 / 偏离 / 说明」这些列）。**不是用来逐条响应、只是要求提交的技术文件**（例如评分要的「所投产品技术规格」表），加 `"mode": "presence"`。商务要求要逐项响应但格式章没给表的，可以自拟一张。 + - 有固定行的表用 `fixed_rows`，响应栏按规矩填。 + - 资格证明、评分证明材料放进**一个** `kind: "attachments"` 的表单（它会把全部附件渲染一遍）。 +4. `attachments`：规矩投标人要附的每份证明材料（按本文件要求来；写明「投标文件中无需提供」的、中标后才交的不附）。证书类 `scan: true`，`lines` 写虚构内容（单位名 {bidder}，信用代码 {credit_code}，有效期覆盖截止日之后至少一年，社保月份按要求的区间）。全部虚构，不出现真实企业或真实人名。 +5. `routes`：tsv 里**每一个** iid 都要分配： + - `tech_row`：技术参数、功能要求；`commercial_row`：商务要求（交货期、地点、质保、付款、售后响应等）； + - `attach:<附件 key>`：靠附证明材料满足的；`form:<表单 key>`：靠某份格式文件满足的； + - `prose` 或 `prose:<节标题>`：靠技术方案散文响应的（实施方案、售后服务方案等评分项）； + - 一条要求要走几处时用 `+` 连起来，例如 `tech_row+attach:a3`（表里一行，另附证明材料）； + - `na`：不需要投标人响应的（无效情形的列举、评审程序、日期、采购人义务、合同条款、说明标注规则的句子、别的包的条目）。 +6. `prose`：2～4 节虚构的技术方案叙述，每节 2～4 段，节标题要和 `prose:<节标题>` 对得上。**不写任何具体参数数值**，不出现别的项目名称。 + +## 自查（用 Python） + +JSON 能解析；tsv 里的 iid 全部出现在 routes 里、没有多余的键；`form:` / `attach:` / `prose:` 都指向存在的表单、附件、节；modes 的 basis 片段都能在对应页逐字找到。可以把 build.py 的输出目录临时指到 scratchpad 试跑一次（不要写进仓库）。 + +## 最终回复 + +只写简短汇总：routes 各取值的条数、forms 列表（key + kind + 有无 mode）、modes 列表、拿不准的地方。 diff --git a/evals/check/prompts/style-B.md b/evals/check/prompts/style-B.md new file mode 100644 index 0000000..e771b2d --- /dev/null +++ b/evals/check/prompts/style-B.md @@ -0,0 +1,27 @@ +# 风格 B「简写型」改写说明（开发集用） + +你在为评测集改写虚构投标文件的技术响应表。输入是一组表格行，每行有编号 `key`、招标要求原文 `requirement`、规矩的响应 `answer`。按下面的风格改写，**只改写法，不改内容**： + +## 改写 `requirement`（招标要求那一栏） + +真实投标人常常不照抄要求原文。每行按下面几种做法之一改写（轮换使用，大约各占三分之一）： + +- **截断**：只留前半句或关键词，后面用「……」或直接截掉。例：「支持多种接入方式，包括有线、无线和蜂窝网络，并支持自动切换」→「支持多种接入方式……」 +- **改写**：换一种说法概括。例：「最小报文发送间隔不大于 50 毫秒」→「报文间隔≤50ms」 +- **去掉编号和标记**：删掉原文开头的序号、★、▲。 + +## 改写 `answer`（投标响应那一栏） + +- 大约三分之一的行**只写值、不复述要求**：例「64GB DDR4」「支持」「3 年」。这些行只输出 `answer`，不输出 `requirement`。 +- 其余行用简短的说法，例「满足，单机 64GB」「支持，含硬件 BFD」。 +- **数值、单位、是否满足必须和原响应完全一致**：原响应写 64GB，改写后也必须是 64GB（不能写成 64G 或 65536MB）；原响应是满足，改写后不能出现「不支持」「部分」「可选」这类字眼。脚本会核对数字，对不上的改写会被丢弃。 + +## 不要做的 + +- 不要编造原文没有的参数、型号、品牌。 +- 不要改变响应的含义（满足的仍然满足）。 +- 例子里的设备和参数是虚构的，不要照搬到输出里。 + +## 输出 + +一个 JSON 对象：`{"": {"requirement": "…", "answer": "…"}, …}`，只写 `answer` 的行省略 `requirement`。每个输入行都要有一项。