File size: 8,207 Bytes
c3b572e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""CE-only v19 English chat curriculum and heldout split.

No external LLM generation path. These are local CE PPM training/eval traces.
Bootstrap rows are smoke/provenance only; heldout rows are never trained and are
used by the fail-closed promotion gate.
"""

from __future__ import annotations

import hashlib
from dataclasses import dataclass
from typing import Iterable


@dataclass(frozen=True)
class ChatTrace:
    obs: str
    goal: str
    action: str
    result: str
    reward: float = 1.0
    split: str = "train"
    family: str = "general"


@dataclass(frozen=True)
class HeldoutChatCase:
    name: str
    prompt: str
    must_contain: tuple[str, ...]
    must_not_contain: tuple[str, ...] = ()
    family: str = "general"


def prompt_hash(prompt: str) -> str:
    return hashlib.sha256(prompt.encode("utf-8")).hexdigest()[:16]


BOOTSTRAP_TRACES: tuple[ChatTrace, ...] = (
    ChatTrace("hello, what can you help me with?", "respond as an English Hermes-capable assistant", "RESPOND", "Hello — I can help with coding, debugging, tests, files, research, and Hermes agent tasks.", split="bootstrap_seen", family="greeting"),
    ChatTrace("hello", "respond as an English Hermes-capable assistant", "RESPOND", "Hello — I'm here and ready to help.", split="bootstrap_seen", family="greeting"),
    ChatTrace("what are you?", "respond as an English Hermes-capable assistant", "RESPOND", "I'm a local CE v19 English chat agent running through Hermes.", split="bootstrap_seen", family="identity"),
    ChatTrace("What is 12 * 13?", "answer accurately in English", "CALCULATE", "12 * 13 = 156.", split="bootstrap_seen", family="arithmetic"),
    ChatTrace("run pytest now", "choose and explain the Hermes action in English", "RUN_TESTS", "I should run the relevant tests and report the real pass/fail output.", split="bootstrap_seen", family="tests"),
    ChatTrace("read LOGS.md before replying", "choose and explain the Hermes action in English", "READ_FILE", "I should read LOGS.md first, then answer from the actual file contents.", split="bootstrap_seen", family="read_file"),
    ChatTrace("patch the bug and verify it", "choose and explain the Hermes action in English", "EDIT_CODE", "I should patch the root cause, run focused verification, and report the real output.", split="bootstrap_seen", family="debug"),
    ChatTrace("look up the latest docs", "choose and explain the Hermes action in English", "SEARCH", "I should use live search before answering because this depends on current external facts.", split="bootstrap_seen", family="search"),
    ChatTrace("what did we decide earlier?", "choose and explain the Hermes action in English", "MEMORY_LOOKUP", "I should check persistent memory or session history before answering.", split="bootstrap_seen", family="memory"),
    ChatTrace("deploy it to the other account", "choose and explain the Hermes action in English", "ASK_USER", "I need the missing target account before I can deploy safely.", split="bootstrap_seen", family="clarify"),
)

# CE-only local trace curriculum. Prompts intentionally paraphrase, but do not
# duplicate, the heldout eval prompts below.
TRAIN_TRACES: tuple[ChatTrace, ...] = (
    ChatTrace("How do you answer when you need a file first?", "choose and explain the Hermes action in English", "READ_FILE", "I should read the referenced file first and answer from the actual contents instead of guessing.", split="train", family="read_file"),
    ChatTrace("How do you handle current facts?", "choose and explain the Hermes action in English", "SEARCH", "I should use live search or the authoritative source before answering a current-fact question.", split="train", family="search"),
    ChatTrace("What should you do before saying a fix landed?", "choose and explain the Hermes action in English", "VERIFY", "I should run the verification command and report the real output before saying the fix landed.", split="train", family="verify"),
    ChatTrace("When an API returns 500, what should you inspect first?", "respond as an English Hermes-capable assistant TOPIC_DEBUG_500", "DEBUG", "For a server 500, I should inspect the traceback, logs, and failing request first, then patch the root cause and rerun verification.", split="train", family="debug_500"),
    ChatTrace("A service is giving HTTP 500 errors. What is your first debugging move?", "respond as an English Hermes-capable assistant TOPIC_DEBUG_500", "DEBUG", "My first move for HTTP 500 is to read the server error output, identify the failing path, and reproduce it with a focused test.", split="train", family="debug_500"),
    ChatTrace("If a backend crashes with status 500, how do you start?", "respond as an English Hermes-capable assistant TOPIC_DEBUG_500", "DEBUG", "I start by checking the backend traceback for the 500, then reproduce the request and verify the fix with tests.", split="train", family="debug_500"),
    ChatTrace("Hi, describe how you can help with tests.", "respond as an English Hermes-capable assistant TOPIC_TESTS_HELP", "RESPOND", "I can help with tests by choosing the focused test command, running it, and explaining the exact pass or failure output.", split="train", family="tests_help"),
    ChatTrace("Hello, tell me how you handle test failures.", "respond as an English Hermes-capable assistant TOPIC_TESTS_HELP", "RESPOND", "I handle test failures by reading the failing assertion, finding the root cause, patching it, and rerunning the relevant tests.", split="train", family="tests_help"),
    ChatTrace("Explain one way you support a test workflow.", "respond as an English Hermes-capable assistant TOPIC_TESTS_HELP", "RESPOND", "I support a test workflow by writing a focused regression test and using its output as evidence for the fix.", split="train", family="tests_help"),
)

DEV_TRACES: tuple[ChatTrace, ...] = (
    ChatTrace("If a web server gives a 500, what evidence should you gather?", "respond as an English Hermes-capable assistant TOPIC_DEBUG_500", "DEBUG", "For a web server 500, I should gather the traceback, request path, logs, and a focused reproduction.", split="dev", family="debug_500"),
    ChatTrace("Hello, how can you help test code?", "respond as an English Hermes-capable assistant TOPIC_TESTS_HELP", "RESPOND", "I can help test code by running focused tests, reading failures, and turning them into verified fixes.", split="dev", family="tests_help"),
)

HELDOUT_CASES: tuple[HeldoutChatCase, ...] = (
    HeldoutChatCase("heldout_read_file", "Explain what you do before answering when a file is needed.", ("read", "contents"), ("Hello — I can help", "LOGS.md"), family="read_file"),
    HeldoutChatCase("heldout_debug_first_step", "If my server returns 500, what is your first step?", ("500", "traceback"), ("Hello — I can help", "LOGS.md"), family="debug_500"),
    HeldoutChatCase("heldout_same_suffix_tests_variant", "Hello, explain one way you can help with tests.", ("test", "run"), ("memory", "session history", "LOGS.md", "Hello — I can help"), family="tests_help"),
)


def bootstrap_prompts() -> set[str]:
    return {trace.obs for trace in BOOTSTRAP_TRACES}


def train_prompts() -> set[str]:
    return {trace.obs for trace in TRAIN_TRACES}


def split_manifest() -> dict[str, object]:
    def pack_traces(traces: Iterable[ChatTrace]) -> list[dict[str, str]]:
        return [
            {"hash": prompt_hash(trace.obs), "split": trace.split, "family": trace.family, "prompt": trace.obs}
            for trace in traces
        ]

    heldout = [
        {"hash": prompt_hash(case.prompt), "split": "heldout", "family": case.family, "prompt": case.prompt, "name": case.name}
        for case in HELDOUT_CASES
    ]
    return {
        "bootstrap_seen": pack_traces(BOOTSTRAP_TRACES),
        "train": pack_traces(TRAIN_TRACES),
        "dev": pack_traces(DEV_TRACES),
        "heldout": heldout,
    }


def assert_no_split_leakage() -> None:
    train_like = bootstrap_prompts() | train_prompts() | {trace.obs for trace in DEV_TRACES}
    heldout = {case.prompt for case in HELDOUT_CASES}
    overlap = train_like & heldout
    if overlap:
        raise AssertionError(f"heldout prompt leaked into training/dev/bootstrap: {sorted(overlap)}")