File size: 8,453 Bytes
b630916
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6b7050a
b630916
 
 
 
6b7050a
 
 
 
 
 
 
 
 
b630916
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
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
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
import json
import hashlib
from flask import Blueprint, request, jsonify
from datasets import load_dataset

bp = Blueprint("rlm_eval_datasets", __name__, url_prefix="/api/rlm-eval/datasets")

_cache: dict[str, dict] = {}


def _make_id(repo: str, config: str, split: str) -> str:
    key = f"{repo}:{config}:{split}"
    return hashlib.md5(key.encode()).hexdigest()[:12]


def _build_hierarchy(rows: list[dict]) -> dict:
    """Reconstruct hierarchy from flat rows: examples -> iterations."""
    examples: dict[int, dict] = {}

    for row in rows:
        ei = row.get("example_idx", 0)
        ri = row.get("rlm_iter", 0)

        if ei not in examples:
            examples[ei] = {
                "example_idx": ei,
                "question_text": row.get("question_text", ""),
                "eval_correct": row.get("eval_correct"),
                "iterations": {},
                "total_input_tokens": 0,
                "total_output_tokens": 0,
                "total_execution_time": 0.0,
                "final_answer": None,
                "final_answer_preview": "",
            }

        ex = examples[ei]

        # Parse code blocks, flattening result.stdout -> stdout
        code_blocks = []
        cbj = row.get("code_blocks_json", "")
        if cbj and cbj != "[]":
            try:
                raw_blocks = json.loads(cbj) if isinstance(cbj, str) else cbj
                for cb in raw_blocks:
                    block = {"code": cb.get("code", "")}
                    result = cb.get("result", {})
                    if isinstance(result, dict) and result.get("stdout"):
                        block["stdout"] = result["stdout"]
                    elif cb.get("stdout"):
                        block["stdout"] = cb["stdout"]
                    code_blocks.append(block)
            except (json.JSONDecodeError, TypeError):
                code_blocks = []

        iteration = {
            "rlm_iter": ri,
            "prompt": row.get("prompt", ""),
            "response": row.get("response", ""),
            "model": row.get("model", ""),
            "input_tokens": row.get("input_tokens", 0),
            "output_tokens": row.get("output_tokens", 0),
            "execution_time": row.get("execution_time", 0.0),
            "has_code_blocks": row.get("has_code_blocks", False),
            "code_blocks": code_blocks,
            "final_answer": row.get("final_answer"),
            "timestamp": row.get("timestamp", ""),
        }

        ex["iterations"][ri] = iteration
        ex["total_input_tokens"] += iteration["input_tokens"] or 0
        ex["total_output_tokens"] += iteration["output_tokens"] or 0
        ex["total_execution_time"] += iteration["execution_time"] or 0.0

        if iteration["final_answer"]:
            ex["final_answer"] = iteration["final_answer"]
            ex["final_answer_preview"] = (iteration["final_answer"] or "")[:200]

    # Sort and convert dicts to lists
    result = []
    for ei_key in sorted(examples.keys()):
        ex = examples[ei_key]
        iters_list = []
        for ri_key in sorted(ex["iterations"].keys()):
            iters_list.append(ex["iterations"][ri_key])
        ex["iterations"] = iters_list
        result.append(ex)

    return {"examples": result}


@bp.route("/load", methods=["POST"])
def load_dataset_endpoint():
    data = request.get_json()
    repo = data.get("repo", "").strip()
    if not repo:
        return jsonify({"error": "repo is required"}), 400

    config = data.get("config", "rlm_call_traces")
    split = data.get("split", "train")

    try:
        ds = load_dataset(repo, config, split=split)
    except Exception as e:
        return jsonify({"error": f"Failed to load dataset: {e}"}), 400

    ds_id = _make_id(repo, config, split)
    rows = [ds[i] for i in range(len(ds))]
    hierarchy = _build_hierarchy(rows)

    # Extract metadata from first row
    first_row = rows[0] if rows else {}
    metadata = {
        "run_id": first_row.get("run_id", ""),
        "method": first_row.get("method", ""),
        "model": first_row.get("model", ""),
    }

    _cache[ds_id] = {
        "repo": repo,
        "config": config,
        "split": split,
        "hierarchy": hierarchy,
        "metadata": metadata,
        "n_rows": len(rows),
    }

    short_name = repo.rsplit("/", 1)[-1] if "/" in repo else repo

    return jsonify({
        "id": ds_id,
        "repo": repo,
        "name": short_name,
        "config": config,
        "split": split,
        "metadata": metadata,
        "n_examples": len(hierarchy["examples"]),
        "n_rows": len(rows),
    })


@bp.route("/", methods=["GET"])
def list_datasets():
    result = []
    for ds_id, info in _cache.items():
        result.append({
            "id": ds_id,
            "repo": info["repo"],
            "name": info["repo"].rsplit("/", 1)[-1] if "/" in info["repo"] else info["repo"],
            "config": info["config"],
            "split": info["split"],
            "metadata": info["metadata"],
            "n_rows": info["n_rows"],
            "n_examples": len(info["hierarchy"]["examples"]),
        })
    return jsonify(result)


@bp.route("/<ds_id>/overview", methods=["GET"])
def get_overview(ds_id):
    """Level 1: Summary of all examples."""
    if ds_id not in _cache:
        return jsonify({"error": "Dataset not loaded"}), 404

    info = _cache[ds_id]
    hierarchy = info["hierarchy"]

    summaries = []
    for ex in hierarchy["examples"]:
        summaries.append({
            "example_idx": ex["example_idx"],
            "question_text": (ex["question_text"] or "")[:300],
            "eval_correct": ex["eval_correct"],
            "n_iterations": len(ex["iterations"]),
            "total_input_tokens": ex["total_input_tokens"],
            "total_output_tokens": ex["total_output_tokens"],
            "total_execution_time": ex["total_execution_time"],
            "final_answer_preview": ex["final_answer_preview"],
        })

    return jsonify({
        "metadata": info["metadata"],
        "examples": summaries,
    })


@bp.route("/<ds_id>/example/<int:example_idx>", methods=["GET"])
def get_example_detail(ds_id, example_idx):
    """Level 2: Iteration timeline for one example."""
    if ds_id not in _cache:
        return jsonify({"error": "Dataset not loaded"}), 404

    info = _cache[ds_id]
    hierarchy = info["hierarchy"]

    ex_data = None
    for ex in hierarchy["examples"]:
        if ex["example_idx"] == example_idx:
            ex_data = ex
            break

    if ex_data is None:
        return jsonify({"error": f"Example {example_idx} not found"}), 404

    iters = []
    for it in ex_data["iterations"]:
        iters.append({
            "rlm_iter": it["rlm_iter"],
            "model": it["model"],
            "input_tokens": it["input_tokens"],
            "output_tokens": it["output_tokens"],
            "execution_time": it["execution_time"],
            "has_code_blocks": it["has_code_blocks"],
            "n_code_blocks": len(it["code_blocks"]),
            "response_preview": (it["response"] or "")[:300],
            "has_final_answer": it["final_answer"] is not None,
            "timestamp": it["timestamp"],
        })

    return jsonify({
        "example_idx": example_idx,
        "question_text": ex_data["question_text"],
        "eval_correct": ex_data["eval_correct"],
        "total_input_tokens": ex_data["total_input_tokens"],
        "total_output_tokens": ex_data["total_output_tokens"],
        "total_execution_time": ex_data["total_execution_time"],
        "final_answer": ex_data["final_answer"],
        "iterations": iters,
    })


@bp.route("/<ds_id>/example/<int:example_idx>/iter/<int:rlm_iter>", methods=["GET"])
def get_iter_detail(ds_id, example_idx, rlm_iter):
    """Full detail for a specific RLM iteration within an example."""
    if ds_id not in _cache:
        return jsonify({"error": "Dataset not loaded"}), 404

    info = _cache[ds_id]
    hierarchy = info["hierarchy"]

    for ex in hierarchy["examples"]:
        if ex["example_idx"] != example_idx:
            continue
        for it in ex["iterations"]:
            if it["rlm_iter"] == rlm_iter:
                return jsonify(it)

    return jsonify({"error": "Iteration not found"}), 404


@bp.route("/<ds_id>", methods=["DELETE"])
def unload_dataset(ds_id):
    if ds_id in _cache:
        del _cache[ds_id]
    return jsonify({"status": "ok"})