"""CBX RAG evaluator v0.1: four Ettin-68M judges for retrieval-augmented answers. from cbx_rag_eval import Faithfulness, ResponseRelevance, ContextRelevance, ContextUtilization judge = ContextRelevance(".") # folder with model.safetensors + cbx_rag_config.json out = judge(question, passages) # passages: str or list of str (or list of sentence lists) Every judge works on sentences. Raw text is split per profile (`split_sentences`, chosen on validation): "contract" splits on line breaks and then with NLTK punkt; "financial" splits on blank lines and then with pysbd. Pass pre-split sentences (a list of strings per passage, `answer_sentences=`) to skip the splitter. Profiles (settings fixed on validation, see the model card): "contract" long single documents (CUAD-like) "financial" several short passages, numbers and tables (FinQA-like); faithfulness adds a numeric check "auto" (default) "contract" for a single passage, "financial" for two or more """ import json import math import os import re from collections import Counter import numpy as np import torch import torch.nn as nn from safetensors.torch import load_file from transformers import AutoConfig, AutoModel, AutoTokenizer try: from cbx_rag_numcheck import Checker except ImportError: # development tree from numcheck import Checker VERSION = "v0.1" # ---------------------------------------------------------------- sentences _WS = re.compile(r"\s+") _LINE = re.compile(r"\n+") def _punkt(): import nltk try: nltk.data.find("tokenizers/punkt_tab") except LookupError: nltk.download("punkt_tab", quiet=True) from nltk.tokenize import sent_tokenize return sent_tokenize _PARA = re.compile(r"\n\s*\n") _SEG = None def _pysbd(): global _SEG if _SEG is None: import warnings with warnings.catch_warnings(): warnings.simplefilter("ignore") import pysbd _SEG = pysbd.Segmenter(language="en", clean=False) return _SEG def split_sentences(text, profile="contract"): """Whitespace-collapsed sentences. "contract": line breaks, then NLTK punkt inside each line. "financial": blank lines, then pysbd inside each paragraph. A JSON table ('[[' ...) stays one sentence.""" text = str(text) if text.lstrip().startswith("[["): return [text.strip()] if profile == "financial": parts = [x for p in _PARA.split(text) for x in _pysbd().segment(p)] else: tok = _punkt() parts = [x for line in _LINE.split(text) for x in tok(line)] out = [_WS.sub(" ", x).strip() for x in parts] return [x for x in out if x] def _letters(n): s = "" n += 1 while n: n, r = divmod(n - 1, 26) s = chr(97 + r) + s return s def _passages(passages): return [passages] if isinstance(passages, str) else list(passages) def _context(passages, profile): """passages: str | list of str | list of list of str -> list of {key, doc, text}.""" ctx = [] for di, p in enumerate(_passages(passages)): sents = list(p) if isinstance(p, (list, tuple)) else split_sentences(p, profile) ctx += [{"key": f"{di}{_letters(j)}", "doc": di, "text": str(s)} for j, s in enumerate(sents)] return ctx def _answer(answer, answer_sentences, profile): sents = list(answer_sentences) if answer_sentences is not None else split_sentences(answer, profile) return [{"key": _letters(j), "text": str(s)} for j, s in enumerate(sents)] def _profile(profile, passages): """"auto": one passage -> "contract", two or more -> "financial" (fixed on validation).""" if profile in ("contract", "financial"): return profile if profile not in (None, "auto"): raise ValueError("profile must be 'auto', 'contract' or 'financial'") return "contract" if len(_passages(passages)) <= 1 else "financial" # ---------------------------------------------------------------- BM25 (as in training) _TOKEN = re.compile(r"[a-z0-9]+") _STOP = set("the a an of to and or in on for by with is are be this that any as at from it its such shall may " "will which their there other than not no does do".split()) def _toks(text): return [t for t in _TOKEN.findall(text.lower()) if t not in _STOP] class BM25: def __init__(self, docs, k1=1.2, b=0.75): self.docs = [Counter(_toks(d)) for d in docs] self.len = [sum(d.values()) for d in self.docs] self.avg = (sum(self.len) / len(self.len)) if self.len else 1 df = Counter() for d in self.docs: df.update(d.keys()) n = len(docs) self.idf = {t: math.log(1 + (n - c + 0.5) / (c + 0.5)) for t, c in df.items()} self.k1, self.b = k1, b def scores(self, query): q = Counter(_toks(query)) out = [] for d, ln in zip(self.docs, self.len): s = 0.0 for t in q: if t in d: tf = d[t] s += self.idf[t] * tf * (self.k1 + 1) / (tf + self.k1 * (1 - self.b + self.b * ln / self.avg)) out.append(s) return out # ---------------------------------------------------------------- model class _Classifier(nn.Module): def __init__(self, config, n_classes): super().__init__() self.backbone = AutoModel.from_config(config, attn_implementation="sdpa") self.head = nn.Linear(config.hidden_size, n_classes) def forward(self, input_ids, attention_mask): h = self.backbone(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state m = attention_mask.unsqueeze(-1).to(h.dtype) return self.head((h * m).sum(1) / m.sum(1).clamp(min=1)) class _Judge: TASK = None def __init__(self, path=".", device=None, batch_tokens=None): with open(os.path.join(path, "cbx_rag_config.json")) as f: self.cfg = json.load(f) if self.cfg["task"] != self.TASK: raise ValueError(f"{path} holds the {self.cfg['task']} model, not {self.TASK}") self.device = device or ("cuda" if torch.cuda.is_available() else "cpu") self.tok = AutoTokenizer.from_pretrained(path) self.model = _Classifier(AutoConfig.from_pretrained(path), self.cfg["n_classes"]) self.model.load_state_dict(load_file(os.path.join(path, "model.safetensors"))) self.model.to(self.device).eval() self.batch_tokens = batch_tokens or self.cfg["batch_tokens"] self._len = {} def _ids(self, text): return self.tok(text, add_special_tokens=False)["input_ids"] def _encode(self, rows): cls, sep = self.tok.cls_token_id, self.tok.sep_token_id out = [] for q, ev, s in rows: qi = self._ids(q)[:self.cfg["q_max"]] si = self._ids(s)[:384] ei = self._ids(ev) room = self.cfg["max_len"] - len(qi) - len(si) - 4 out.append([cls] + qi + [sep] + ei[:max(0, room)] + [sep] + si + [sep]) return out @torch.no_grad() def _probs(self, rows): """rows: (question, evidence, sentence) -> class probabilities [n, n_classes].""" enc = self._encode(rows) probs = np.zeros((len(enc), self.cfg["n_classes"]), dtype=np.float32) if not enc: return probs lengths = [len(x) for x in enc] order = sorted(range(len(enc)), key=lambda i: lengths[i]) pad = self.tok.pad_token_id cur = [] def flush(): n = max(lengths[i] for i in cur) ids = torch.full((len(cur), n), pad, dtype=torch.long) mask = torch.zeros((len(cur), n), dtype=torch.long) for k, i in enumerate(cur): ids[k, :lengths[i]] = torch.tensor(enc[i]) mask[k, :lengths[i]] = 1 ids, mask = ids.to(self.device), mask.to(self.device) if self.device.startswith("cuda"): with torch.autocast("cuda", dtype=torch.float16): logits = self.model(ids, mask).float() else: logits = self.model(ids, mask).float() probs[cur] = torch.softmax(logits, -1).cpu().numpy() for i in order: if cur and lengths[i] * (len(cur) + 1) > self.batch_tokens * 2: flush() cur = [] cur.append(i) flush() return probs # ---------------------------------------------------------------- faithfulness class Faithfulness(_Judge): """Is each answer sentence supported by the passages? Returns sentence scores, an answer score and the hallucination flag (an answer is flagged when any sentence is).""" TASK = "faithfulness" VERDICTS = ("all_verified", "some_unverified") def _ntok(self, text): n = self._len.get(text) if n is None: n = len(self._ids(text)) if len(self._len) < 200_000: self._len[text] = n return n def _line(self, c): text = c["text"] cap = self.cfg["sentence_cap_tokens"] if self._ntok(text) > cap: text = self.tok.decode(self._ids(text)[:cap]) + " ..." return f"[{c['key']}] {text}" def _evidence(self, question, ctx, sents, si, bm25): budget = self.cfg["evidence_budget"] overhead = self._ntok(question) + self._ntok(sents[si]["text"]) + 8 room = max(64, budget - overhead) lines = [self._line(c) for c in ctx] lens = [self._ntok(x) for x in lines] if sum(lens) <= room: return "\n".join(lines) chosen, used = set(), 0 def take(i): nonlocal used if i in chosen or i < 0 or i >= len(ctx) or used + lens[i] > room: return False chosen.add(i) used += lens[i] return True query = question + " " + sents[si]["text"] if si > 0: query += " " + sents[si - 1]["text"] sc = bm25.scores(query) for i in sorted(range(len(ctx)), key=lambda i: -sc[i]): if used >= room - 16 or sc[i] <= 0: break for j in (i, i - 1, i + 1): take(j) if used < room: for i in range(len(ctx)): if not take(i) and used + min(lens) > room: break return "\n".join(lines[i] for i in sorted(chosen)) def __call__(self, question, passages, answer, answer_sentences=None, profile="auto", operating_point="default"): """operating_point: "default" (flags about 4.5% of sentences on validation) or "high_recall".""" prof = _profile(profile, passages) ctx = _context(passages, prof) sents = _answer(answer, answer_sentences, prof) pc = self.cfg["profiles"][prof] bm25 = BM25([c["text"] for c in ctx]) rows = [(question, self._evidence(question, ctx, sents, si, bm25), s["text"]) for si, s in enumerate(sents)] p = 1 - self._probs(rows)[:, 0] # probability "unsupported" score, checks = p.astype(float), [None] * len(sents) thr = pc["thresholds"][operating_point] if pc.get("numeric_check") and operating_point == "default": ck = Checker(" ".join(c["text"] for c in ctx)) checks = ck.check_answer([s["text"] for s in sents]) c = pc["combiner"] X = [] for pi, nc in zip(p, checks): q = min(max(float(pi), 1e-6), 1 - 1e-6) X.append([math.log(q / (1 - q))] + [float(nc["verdict"] == v) for v in self.VERDICTS] + [min(nc["n_unverified"], 3), float(nc["eq_bad"] > 0), float(nc["eq_ok"] > 0)]) score = 1 / (1 + np.exp(-(np.array(X) @ np.array(c["coef"]) + c["intercept"]))) out = [{"key": s["key"], "text": s["text"], "unsupported_prob": round(float(pi), 6), "score": round(float(si), 6), "flag": bool(si >= thr), **({"numbers": checks[k]["verdict"]} if checks[k] else {})} for k, (s, pi, si) in enumerate(zip(sents, p, score))] return {"profile": prof, "operating_point": operating_point, "sentences": out, "answer_score": round(float(max(score)) if len(out) else 0.0, 6), "hallucination": any(x["flag"] for x in out)} # ---------------------------------------------------------------- response relevance class ResponseRelevance(_Judge): """Does the answer address the question? direct / partial / irrelevant (question and answer only).""" TASK = "response_relevance" def __call__(self, question, answer): return self.batch([(question, answer)])[0] def batch(self, pairs): pr = self._probs([(q, "", a) for q, a in pairs]) labels = self.cfg["labels"] return [{"label": labels[int(x.argmax())], "probs": {l: round(float(v), 6) for l, v in zip(labels, x)}} for x in pr] # ---------------------------------------------------------------- context relevance / utilization def _answer_score(p, spec): if spec["agg"] == "count": return float((p >= spec["thr"]).mean()) if spec["agg"] == "mean_iso": return float(np.interp(p, spec["iso_x"], spec["iso_y"]).mean()) return float(p.mean()) def _ap_at_k(v): v = np.asarray(v, float) if v.sum() == 0: return 0.0 return float(((np.cumsum(v) / np.arange(1, len(v) + 1)) * v).sum() / v.sum()) class _ContextJudge(_Judge): def _rows(self, q, ctx, idx): w, ch = self.cfg["window"], self.cfg["neighbour_chars"] rows = [] for i in idx: c = ctx[i] js = [j for j in range(i - w, i + w + 1) if j != i and 0 <= j < len(ctx) and ctx[j]["doc"] == c["doc"]] rows.append((q, " ".join(ctx[j]["text"][:ch] for j in js), c["text"])) return rows def _run(self, q, ctx, pc, prefilter_query=None): idx = list(range(len(ctx))) top = pc.get("prefilter_top_n") if top and prefilter_query is not None and len(ctx) > top: sc = np.asarray(BM25([c["text"] for c in ctx]).scores(prefilter_query)) m = np.zeros(len(ctx), bool) for i in np.argsort(-sc)[:top]: m[max(0, i - 1):i + 2] = True idx = [i for i in idx if m[i]] p = np.zeros(len(ctx)) if idx: p[idx] = 1 - self._probs(self._rows(q, ctx, idx))[:, 0] return p, len(idx) class ContextRelevance(_ContextJudge): """Is each context sentence relevant to the question? Returns sentence scores, the answer-level relevance fraction and, with two or more passages, context precision (AP over passages in the given order).""" TASK = "context_relevance" def __call__(self, question, passages, profile="auto"): prof = _profile(profile, passages) ctx = _context(passages, prof) pc = self.cfg["profiles"][prof] p, _ = self._run(question, ctx, pc) thr = pc["sentence_threshold"] out = {"profile": prof, "sentences": [{"key": c["key"], "passage": c["doc"], "text": c["text"], "score": round(float(x), 6), "relevant": bool(x >= thr)} for c, x in zip(ctx, p)], "relevance": round(_answer_score(p, pc["answer_score"]), 6)} docs = sorted({c["doc"] for c in ctx}) if len(docs) > 1: d = np.array([c["doc"] for c in ctx]) out["passage_relevant"] = [bool((p[d == k] >= thr).any()) for k in docs] out["context_precision"] = round(_ap_at_k(out["passage_relevant"]), 6) return out class ContextUtilization(_ContextJudge): """Did the answer use each context sentence? Returns sentence scores and the utilization fraction.""" TASK = "context_utilization" def __call__(self, question, passages, answer, profile="auto", prefilter=True): prof = _profile(profile, passages) ctx = _context(passages, prof) pc = dict(self.cfg["profiles"][prof]) if not prefilter: pc["prefilter_top_n"] = None q = question + " || " + str(answer)[:1200] p, n_scored = self._run(q, ctx, pc, prefilter_query=question + " " + str(answer)) thr = pc["sentence_threshold"] return {"profile": prof, "sentences_scored": n_scored, "sentences": [{"key": c["key"], "passage": c["doc"], "text": c["text"], "score": round(float(x), 6), "utilized": bool(x >= thr)} for c, x in zip(ctx, p)], "utilization": round(_answer_score(p, pc["answer_score"]), 6)}