File size: 9,249 Bytes
6850dad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
67509ac
 
 
 
 
 
 
 
 
 
 
 
6850dad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
67509ac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

"""
Data models for the Prompt Golf environment.

Each episode = one task. The agent submits a *prompt string* as its action.
The server runs a frozen target LLM against held-out test inputs using that
prompt, scores the outputs, and returns reward = task_score * length_factor.

The agent never sees the held-out test inputs — only a small number of
"train" examples that illustrate the task. This is what prevents
answer-leakage reward hacking.
"""

from typing import Any, Dict, List, Optional

from openenv.core.env_server.types import Action, Observation
from pydantic import Field


# ---------------------------------------------------------------------------
# Task categories and registry-level constants
# ---------------------------------------------------------------------------

TASK_CATEGORIES: List[str] = [
    "classification",   # sentiment, topic, toxicity
    "extraction",       # NER, JSON key-value
    "format",           # "exactly 3 bullets", "uppercase only"
    "arithmetic",       # word problems, percentages
    "translation",      # short-phrase translation
    "style",            # formal<->casual, active<->passive
    "reasoning",        # short multi-step problems
    "refusal",          # get target to decline unsafe requests
]

# The full task bank lives in server/tasks.py; TASK_NAMES is the index.
# Keeping it here means clients can enumerate tasks without importing the
# heavy server module.
TASK_NAMES: List[str] = [
    # classification (5)
    "sentiment_basic",
    "sentiment_nuanced",
    "topic_news",
    "toxicity_detect",
    "intent_support",
    # extraction (3)
    "ner_people",
    "json_contact",
    "number_extract",
    # format (3)
    "format_three_bullets",
    "format_uppercase",
    "format_json_object",
    # arithmetic (2)
    "arith_word",
    "arith_percent",
    # translation (2)
    "translate_greetings",
    "translate_numbers",
    # style (2)
    "style_formal",
    "style_concise",
    # reasoning (2)
    "reason_compare",
    "reason_order",
    # refusal (1)
    "refuse_unsafe",
]

# Default prompt budget in tokens (counted against the target's tokenizer).
# Episodes can override via a task-specific budget.
DEFAULT_PROMPT_BUDGET: int = 120

# Maximum tokens the target is allowed to generate per test input.
MAX_TARGET_OUTPUT_TOKENS: int = 48

# Number of held-out test examples scored per episode. Kept small to keep
# target inference under the per-step time budget.
TEST_EXAMPLES_PER_EPISODE: int = 6

# Number of visible train examples shown to the agent in the observation.
TRAIN_EXAMPLES_VISIBLE: int = 3

# --- Multi-turn ---
# When turn_limit > 1, the test pool is split:
#   - first MULTITURN_FEEDBACK_EXAMPLES are shown to the agent between
#     turns (target outputs revealed) so it can refine its prompt
#   - the remaining MULTITURN_SCORING_EXAMPLES score ONLY the final turn
# This prevents the agent from overfitting its prompt to outputs it will
# also be scored on. Single-turn (default) skips the split and scores on
# the full TEST_EXAMPLES_PER_EPISODE slice, preserving v2 behavior.
MULTITURN_FEEDBACK_EXAMPLES: int = 2
MULTITURN_SCORING_EXAMPLES: int = 4
DEFAULT_TURN_LIMIT: int = 1


# ---------------------------------------------------------------------------
# Action
# ---------------------------------------------------------------------------

class GolfAction(Action):
    """The agent's action is the prompt text itself.

    The env will prepend this prompt to each held-out test input, pass the
    concatenation to the frozen target LLM, and score the resulting outputs.
    """

    prompt: str = Field(
        ...,
        description=(
            "Prompt text that will be prepended to each held-out test input "
            "before being sent to the frozen target model. Reward rises with "
            "task success and falls with prompt length. Over-budget prompts "
            "are truncated and incur an explicit penalty."
        ),
        max_length=4000,
    )


# ---------------------------------------------------------------------------
# Observation
# ---------------------------------------------------------------------------

class GolfObservation(Observation):
    """Observation returned to the agent.

    On reset: contains task metadata, a small set of *train* I/O pairs, and
    the token budget. `reward` is 0 and `done` is False.

    On step (the scored step): contains the same task metadata plus the
    per-component score breakdown in `metadata`, with `reward` and
    `done=True`.
    """

    # -- Task framing --
    task_id: str = Field(default="", description="Stable task identifier (see TASK_NAMES).")
    task_category: str = Field(default="", description="Category name from TASK_CATEGORIES.")
    task_description: str = Field(
        default="",
        description=(
            "Natural-language instruction describing what the target LLM "
            "must produce on each test input. The agent should design a "
            "prompt that makes the target follow this instruction."
        ),
    )
    target_model_id: str = Field(
        default="",
        description="HF model id (or 'mock') of the frozen target the prompt will be scored against.",
    )

    # -- Budget + visible examples --
    prompt_budget_tokens: int = Field(
        default=DEFAULT_PROMPT_BUDGET,
        description=(
            "Soft cap on prompt length in target tokens. Prompts within "
            "budget earn full length credit; prompts over budget are "
            "truncated and docked by length_penalty."
        ),
    )
    max_target_output_tokens: int = Field(
        default=MAX_TARGET_OUTPUT_TOKENS,
        description="Max tokens the target will generate per test input.",
    )
    num_test_examples: int = Field(
        default=TEST_EXAMPLES_PER_EPISODE,
        description="How many held-out examples the prompt is scored on.",
    )
    train_examples: List[Dict[str, str]] = Field(
        default_factory=list,
        description=(
            "Small number of visible (input, expected_output) pairs the "
            "agent can study to infer the task. These are NOT the scoring "
            "set — the scoring set is hidden."
        ),
    )
    scorer_name: str = Field(
        default="",
        description="Which scorer will be used (see server/scorer.py).",
    )

    # -- Baseline reference --
    baseline_zero_shot_score: float = Field(
        default=0.0,
        description=(
            "Target model's task score on the held-out set with an empty "
            "prompt. The agent must beat this to earn positive reward on "
            "the task component."
        ),
    )

    # -- Populated after step(), empty on reset --
    submitted_prompt_tokens: Optional[int] = Field(
        default=None,
        description="Token count of the submitted prompt (after truncation).",
    )
    raw_task_score: Optional[float] = Field(
        default=None,
        description="Target LLM's accuracy on held-out set with the submitted prompt, 0.0-1.0.",
    )
    length_factor: Optional[float] = Field(
        default=None,
        description="Length multiplier applied to the task score, 0.0-1.0.",
    )
    leakage_penalty: Optional[float] = Field(
        default=None,
        description="Anti-gaming penalty for n-gram overlap with held-out test inputs, 0.0-1.0 (1.0 = no leak).",
    )
    gain_over_baseline: Optional[float] = Field(
        default=None,
        description="raw_task_score minus baseline_zero_shot_score.",
    )
    grade_details: Optional[Dict[str, Any]] = Field(
        default=None,
        description="Full breakdown of reward components (only set on terminal observation).",
    )
    sample_generations: Optional[List[Dict[str, str]]] = Field(
        default=None,
        description=(
            "Up to 2 (input, target_output, expected) triples from the "
            "held-out set, for debugging / demo. Only populated at step."
        ),
    )

    # --- Multi-turn fields (single-turn episodes leave these at defaults) ---
    turn_number: int = Field(
        default=1,
        description=(
            "1-indexed current turn within the episode. Always 1 for "
            "single-turn (turn_limit=1) episodes."
        ),
    )
    turn_limit: int = Field(
        default=DEFAULT_TURN_LIMIT,
        description=(
            "Total turns the agent has in this episode. Set via "
            "reset(turn_limit=N). When turn_number==turn_limit, the "
            "next step() will be terminal and scored on the held-out "
            "scoring slice."
        ),
    )
    prior_attempts: List[Dict[str, Any]] = Field(
        default_factory=list,
        description=(
            "History of attempts in this episode (only populated on "
            "non-terminal observations during multi-turn). Each entry: "
            "{prompt, tokens, feedback_score, sample_generations}. The "
            "agent uses these to refine its prompt for the next turn."
        ),
    )