eval.dataset_builder

chunktuner.eval.dataset_builder

Build EvalDataset from LLM outputs or user JSON/YAML files.

DatasetBuilder 

DatasetBuilder(llm_model='gpt-4o-mini')

Build EvalDataset from user files or LLM-generated Q&A over documents.

Source code in src/chunktuner/eval/dataset_builder.py 
37
38
39
def __init__(self, llm_model: str = "gpt-4o-mini"):
    self.llm_model = llm_model
    self._enc = tiktoken.get_encoding("cl100k_base")

build_from_user_file 

build_from_user_file(path)

Parse JSON or YAML with a queries list into an EvalDataset (user-provided).

Source code in src/chunktuner/eval/dataset_builder.py 
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
def build_from_user_file(self, path: Path) -> EvalDataset:
    """Parse JSON or YAML with a ``queries`` list into an `EvalDataset` (user-provided)."""
    raw = path.read_text(encoding="utf-8")
    if path.suffix.lower() in {".yaml", ".yml"}:
        data = yaml.safe_load(raw)
    else:
        data = json.loads(raw)
    queries = []
    for row in data.get("queries", []):
        spans = [tuple(x) for x in row.get("answer_spans", [])]
        queries.append(
            EvalQuery(
                id=row.get("id", str(uuid.uuid4())),
                question=row["question"],
                document_id=row["document_id"],
                answer_spans=spans,
                reference_answer=row.get("reference_answer"),
            )
        )
    return EvalDataset(
        name=data.get("name", path.stem),
        queries=queries,
        source="user_provided",
    )

build_from_docs 

build_from_docs(
    docs, *, max_queries=50, questions_per_doc=3
)

LLM-generated Q&A with span validation (requires litellm + API keys).

Source code in src/chunktuner/eval/dataset_builder.py 
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
def build_from_docs(
    self,
    docs: list[Document],
    *,
    max_queries: int = 50,
    questions_per_doc: int = 3,
) -> EvalDataset:
    """LLM-generated Q&A with span validation (requires ``litellm`` + API keys)."""
    import litellm

    if not docs:
        return EvalDataset(name="empty", queries=[], source="llm_generated")
    random.shuffle(docs)
    budget = max(1, max_queries // max(1, questions_per_doc))
    sample = docs[:budget]
    queries: list[EvalQuery] = []
    for d in sample:
        snippet = d.content[:6000]
        prompt = (
            f"Document id: {d.id}\n"
            "Create evaluation questions with answer spans as CHARACTER offsets into the "
            "EXACT document text below. Return JSON: "
            '{"queries":[{"question":str,"start":int,"end":int,"reference_answer":str}]}\n\n'
            f"DOCUMENT:\n{snippet}"
        )
        try:
            resp = litellm.completion(
                model=self.llm_model,
                messages=[{"role": "user", "content": prompt}],
                response_format={"type": "json_object"},
                temperature=0.2,
            )
            raw = resp.choices[0].message.content or "{}"
            payload = json.loads(raw)
        except Exception as exc:
            logger.warning("Dataset generation failed for doc %s: %s", d.id, exc)
            continue
        rows = payload.get("queries", payload) if isinstance(payload, dict) else payload
        if not isinstance(rows, list):
            rows = []
        for row in rows[:questions_per_doc]:
            if not isinstance(row, dict):
                continue
            try:
                q = str(row.get("question", "")).strip()
                a = int(row.get("start", row.get("start_offset", 0)))
                b = int(row.get("end", row.get("end_offset", 0)))
                ref = str(row.get("reference_answer", ""))
            except (TypeError, ValueError) as exc:
                logger.warning("Skipping malformed query row %r: %s", row, exc)
                continue
            span_text = d.content[a:b] if 0 <= a < b <= len(d.content) else ""
            if not q or not span_text:
                continue
            if ref and _token_f1(self._enc, span_text, ref) < 0.85:
                continue
            queries.append(
                EvalQuery(
                    id=f"q_{uuid.uuid4()}",
                    question=q,
                    document_id=d.id,
                    answer_spans=[(a, b)],
                    reference_answer=ref or None,
                )
            )
            if len(queries) >= max_queries:
                break
        if len(queries) >= max_queries:
            break
    if len(queries) < 20:
        logger.warning(
            "Only %d valid eval queries generated (minimum recommended: 20). "
            "Results may be noisy.",
            len(queries),
        )
    return EvalDataset(name="llm_generated", queries=queries, source="llm_generated")

build_code_function_qa 

build_code_function_qa(docs, *, max_queries=40)

Heuristic code questions: point at first function body span per file.

Source code in src/chunktuner/eval/dataset_builder.py 
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
def build_code_function_qa(self, docs: list[Document], *, max_queries: int = 40) -> EvalDataset:
    """Heuristic code questions: point at first function body span per file."""
    queries: list[EvalQuery] = []
    for d in docs:
        if d.content_type != "code":
            continue
        m = re.search(r"(?m)^def\s+\w+\s*\(", d.content)
        if not m:
            continue
        start = m.start()
        end = min(len(d.content), start + 400)
        queries.append(
            EvalQuery(
                id=f"code_{d.id}",
                question="What does the first top-level function in this file do?",
                document_id=d.id,
                answer_spans=[(start, end)],
            )
        )
        if len(queries) >= max_queries:
            break
    return EvalDataset(name="code_heuristic", queries=queries, source="user_provided")