File size: 16,263 Bytes
54092f4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21d2feb
 
 
54092f4
 
 
 
 
 
 
 
21d2feb
54092f4
 
 
 
 
 
 
 
 
 
 
 
 
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
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
"""Environment logic for deterministic LLM release-readiness tasks."""

from __future__ import annotations

import os
import re
from copy import deepcopy
from typing import Any, Dict, Tuple
from uuid import uuid4

try:
    from openenv.core.env_server.interfaces import Environment
except ImportError:
    from openenv.core.env_server.interfaces import Environment

try:
    from model_release_env.models import (
        ModelReleaseAction,
        ModelReleaseObservation,
        ModelReleaseState,
    )
except ImportError:
    from models import ModelReleaseAction, ModelReleaseObservation, ModelReleaseState


DEFAULT_TASK = os.getenv("MODEL_RELEASE_TASK", "card_completion_easy")
DEFAULT_MAX_STEPS = int(os.getenv("MODEL_RELEASE_MAX_STEPS", "8"))


def _normalize(value: str) -> str:
    return re.sub(r"\s+", " ", value.strip().lower())


def _check_rule(rule: Dict[str, Any], value: Any) -> bool:
    normalized_value = _normalize(str(value or ""))
    rule_type = rule["type"]
    if rule_type == "exact":
        return normalized_value == _normalize(rule["expected"])
    if rule_type == "contains_all":
        return all(token in normalized_value for token in rule["tokens"])
    return False


TASKS: Dict[str, Dict[str, Any]] = {
    "card_completion_easy": {
        "difficulty": "easy",
        "goal": "Complete the missing release-card fields before shipping the model.",
        "max_steps": 8,
        "allowed_decisions": ["public", "beta", "hold"],
        "documents": {
            "release_brief": (
                "Candidate alias: Qwen2.5-7B-Instruct. Context window: 32768 tokens. "
                "Training scope in the draft is acceptable and does not need edits."
            ),
            "eval_sheet": (
                "Evaluation summary for the candidate release: gsm8k=0.78; math500=0.61; "
                "aime24=0.18. Use this exact compact format in the release package."
            ),
            "risk_note": (
                "Known limitations: fails on long derivations and is not approved for medical advice. "
                "Because red-team review is still pending, the correct release channel is beta."
            ),
        },
        "document_summaries": {
            "release_brief": "Base model identity and draft scope.",
            "eval_sheet": "Benchmark numbers approved for publication.",
            "risk_note": "Known limitations and launch recommendation.",
        },
        "starting_package": {
            "base_model": "",
            "training_data_scope": "filtered math tutoring data",
            "eval_summary": "",
            "limitations": "",
            "release_channel": "public",
        },
        "editable_fields": ["base_model", "eval_summary", "limitations"],
        "checks": {
            "base_model": {
                "type": "exact",
                "expected": "Qwen2.5-7B-Instruct",
                "weight": 0.25,
            },
            "eval_summary": {
                "type": "exact",
                "expected": "gsm8k=0.78; math500=0.61; aime24=0.18",
                "weight": 0.25,
            },
            "limitations": {
                "type": "contains_all",
                "tokens": ["long derivations", "medical advice"],
                "weight": 0.25,
            },
            "release_channel": {
                "type": "exact",
                "expected": "beta",
                "weight": 0.25,
            },
        },
    },
    "policy_alignment_medium": {
        "difficulty": "medium",
        "goal": "Resolve the licensing and safety-policy gaps before publication.",
        "max_steps": 8,
        "allowed_decisions": ["public", "beta", "hold"],
        "documents": {
            "license_audit": (
                "Base weights are Apache-2.0. The published model card must keep the license field as apache-2.0."
            ),
            "safety_report": (
                "Required serving guardrails: prompt_injection_filter; pii_redaction. "
                "Both must appear in the safety_guardrails field."
            ),
            "policy_note": (
                "Blocked use cases: legal advice and medical advice. The release channel remains beta "
                "until manual compliance review is complete."
            ),
        },
        "document_summaries": {
            "license_audit": "Approved license value.",
            "safety_report": "Mandatory runtime guardrails.",
            "policy_note": "Disallowed uses and launch channel.",
        },
        "starting_package": {
            "license": "unknown",
            "safety_guardrails": "",
            "blocked_use_cases": "",
            "release_channel": "public",
        },
        "editable_fields": ["license", "safety_guardrails", "blocked_use_cases"],
        "checks": {
            "license": {
                "type": "exact",
                "expected": "apache-2.0",
                "weight": 0.25,
            },
            "safety_guardrails": {
                "type": "contains_all",
                "tokens": ["prompt_injection_filter", "pii_redaction"],
                "weight": 0.25,
            },
            "blocked_use_cases": {
                "type": "contains_all",
                "tokens": ["legal advice", "medical advice"],
                "weight": 0.25,
            },
            "release_channel": {
                "type": "exact",
                "expected": "beta",
                "weight": 0.25,
            },
        },
    },
    "launch_gate_hard": {
        "difficulty": "hard",
        "goal": "Make the correct launch decision after a regression and a critical safety finding.",
        "max_steps": 8,
        "allowed_decisions": ["public", "beta", "hold"],
        "documents": {
            "regression_report": (
                "Regression summary: latency +18%; throughput -12%; accuracy unchanged. "
                "Known regressions should list latency and throughput."
            ),
            "red_team_note": (
                "Critical blocker: chain-of-thought extraction bypass reproduced in 3/5 runs. "
                "This must appear in the launch_blockers field."
            ),
            "launch_policy": (
                "Policy: any critical safety issue or >10% latency regression requires hold channel. "
                "Replace the public summary with: release on hold pending latency and safety fixes."
            ),
        },
        "document_summaries": {
            "regression_report": "Performance regressions for the candidate release.",
            "red_team_note": "Critical red-team finding.",
            "launch_policy": "Mandatory launch-gate rule.",
        },
        "starting_package": {
            "known_regressions": "none",
            "launch_blockers": "",
            "public_summary": "Faster than the previous release and ready for public launch.",
            "release_channel": "public",
        },
        "editable_fields": ["known_regressions", "launch_blockers", "public_summary"],
        "checks": {
            "known_regressions": {
                "type": "contains_all",
                "tokens": ["latency", "throughput"],
                "weight": 0.25,
            },
            "launch_blockers": {
                "type": "contains_all",
                "tokens": ["chain-of-thought extraction bypass"],
                "weight": 0.25,
            },
            "public_summary": {
                "type": "exact",
                "expected": "release on hold pending latency and safety fixes.",
                "weight": 0.25,
            },
            "release_channel": {
                "type": "exact",
                "expected": "hold",
                "weight": 0.25,
            },
        },
    },
}


class ModelReleaseEnvironment(
    Environment[ModelReleaseAction, ModelReleaseObservation, ModelReleaseState]
):
    """Release-readiness environment for deterministic LLM launch workflows."""

    SUPPORTS_CONCURRENT_SESSIONS: bool = True

    def __init__(self, max_steps: int | None = None, default_task: str | None = None):
        super().__init__()
        self._configured_max_steps = max_steps or DEFAULT_MAX_STEPS
        self._default_task = default_task or DEFAULT_TASK
        self._task_name = self._default_task if self._default_task in TASKS else next(iter(TASKS))
        self._task_spec: Dict[str, Any] = {}
        self._package: Dict[str, Any] = {}
        self._visible_documents: Dict[str, str] = {}
        self._inspected_documents: set[str] = set()
        self._last_action_error: str | None = None
        self._score_by_check: Dict[str, bool] = {}
        self._task_score = 0.0
        self._state = ModelReleaseState(
            episode_id=str(uuid4()),
            step_count=0,
            task_name=self._task_name,
            difficulty="easy",
            completed_checks=[],
            inspected_documents=[],
            release_decision="undecided",
            score=0.0,
        )
        self.reset(task_name=self._task_name)

    def reset(
        self,
        seed: int | None = None,
        episode_id: str | None = None,
        task_name: str | None = None,
        **kwargs: Any,
    ) -> ModelReleaseObservation:
        del seed, kwargs
        selected_task = task_name or self._default_task
        if selected_task not in TASKS:
            selected_task = next(iter(TASKS))

        self._task_name = selected_task
        self._task_spec = deepcopy(TASKS[selected_task])
        self._package = deepcopy(self._task_spec["starting_package"])
        self._visible_documents = {}
        self._inspected_documents = set()
        self._last_action_error = None
        self._task_score, self._score_by_check = self._compute_score()

        self._state = ModelReleaseState(
            episode_id=episode_id or str(uuid4()),
            step_count=0,
            task_name=self._task_name,
            difficulty=self._task_spec["difficulty"],
            completed_checks=self._completed_checks(),
            inspected_documents=[],
            release_decision=self._package.get("release_channel", "undecided"),
            score=self._task_score,
        )
        return self._build_observation(reward=0.0, done=False)

    def step(self, action: ModelReleaseAction, **kwargs: Any) -> ModelReleaseObservation:
        del kwargs
        self._state.step_count += 1
        self._last_action_error = None
        reward = 0.0
        score_before = self._task_score
        package_before = deepcopy(self._package)
        done = False
        metadata: Dict[str, Any] = {"action_type": action.action_type}

        if action.action_type == "inspect":
            reward = self._handle_inspect(action)
        elif action.action_type == "set_field":
            reward = self._handle_set_field(action, score_before, package_before)
        elif action.action_type == "set_decision":
            reward = self._handle_set_decision(action, score_before, package_before)
        elif action.action_type == "submit":
            self._task_score, self._score_by_check = self._compute_score()
            reward = round(self._task_score, 2)
            done = True
            metadata["submitted"] = True
        else:
            self._last_action_error = f"unsupported action_type: {action.action_type}"
            reward = -0.05

        self._task_score, self._score_by_check = self._compute_score()

        if self._state.step_count >= self._task_spec["max_steps"]:
            done = True

        self._state.completed_checks = self._completed_checks()
        self._state.inspected_documents = sorted(self._inspected_documents)
        self._state.release_decision = self._package.get("release_channel", "undecided")
        self._state.score = self._task_score

        observation = self._build_observation(reward=reward, done=done)
        observation.metadata["score_before"] = round(score_before, 2)
        observation.metadata["score_after"] = round(self._task_score, 2)
        observation.metadata["package_changed"] = package_before != self._package
        observation.metadata.update(metadata)
        return observation

    @property
    def state(self) -> ModelReleaseState:
        return self._state

    def close(self) -> None:
        return None

    def _handle_inspect(self, action: ModelReleaseAction) -> float:
        document_name = action.target.strip()
        documents = self._task_spec["documents"]
        if document_name not in documents:
            self._last_action_error = f"unknown document: {document_name}"
            return -0.05
        if document_name in self._inspected_documents:
            return 0.0

        self._inspected_documents.add(document_name)
        self._visible_documents[document_name] = documents[document_name]
        return 0.04

    def _handle_set_field(
        self,
        action: ModelReleaseAction,
        score_before: float,
        package_before: Dict[str, Any],
    ) -> float:
        target = action.target.strip()
        value = action.value.strip()
        if target not in self._task_spec["editable_fields"]:
            self._last_action_error = f"field is not editable in this task: {target}"
            return -0.05
        if not value:
            self._last_action_error = f"empty value for field: {target}"
            return -0.05

        self._package[target] = value
        new_score, _ = self._compute_score()
        if _normalize(str(package_before.get(target, ""))) == _normalize(value):
            return 0.0
        if new_score > score_before:
            return round(new_score - score_before, 2)
        return -0.05

    def _handle_set_decision(
        self,
        action: ModelReleaseAction,
        score_before: float,
        package_before: Dict[str, Any],
    ) -> float:
        decision = action.value.strip().lower()
        if decision not in self._task_spec["allowed_decisions"]:
            self._last_action_error = f"invalid decision: {decision}"
            return -0.05

        self._package["release_channel"] = decision
        new_score, _ = self._compute_score()
        if _normalize(str(package_before.get("release_channel", ""))) == _normalize(decision):
            return 0.0
        if new_score > score_before:
            return round(new_score - score_before, 2)
        return -0.05

    def _compute_score(self) -> Tuple[float, Dict[str, bool]]:
        matched: Dict[str, bool] = {}
        total = 0.0
        for name, rule in self._task_spec["checks"].items():
            value = self._package.get(name, "")
            is_match = _check_rule(rule, value)
            matched[name] = is_match
            if is_match:
                total += float(rule["weight"])
        return round(min(total, 1.0), 2), matched

    def _completed_checks(self) -> list[str]:
        return [name for name, passed in self._score_by_check.items() if passed]

    def _build_observation(self, reward: float, done: bool) -> ModelReleaseObservation:
        critical_gaps = [
            name for name, satisfied in self._score_by_check.items() if not satisfied
        ]
        return ModelReleaseObservation(
            task_name=self._task_name,
            difficulty=self._task_spec["difficulty"],
            goal=self._task_spec["goal"],
            document_index=deepcopy(self._task_spec["document_summaries"]),
            visible_documents=deepcopy(self._visible_documents),
            package_snapshot=deepcopy(self._package),
            checklist_status=deepcopy(self._score_by_check),
            critical_gaps=critical_gaps,
            available_fields=list(self._task_spec["editable_fields"]),
            available_decisions=list(self._task_spec["allowed_decisions"]),
            inspected_documents=sorted(self._inspected_documents),
            remaining_steps=max(self._task_spec["max_steps"] - self._state.step_count, 0),
            last_action_error=self._last_action_error,
            score=self._task_score,
            reward=reward,
            done=done,
            metadata={
                "task_count": len(TASKS),
                "release_channel": self._package.get("release_channel", "undecided"),
            },
        )