File size: 3,634 Bytes
22decf5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Paths and settings. Everything machine-specific comes from the environment or `.env`."""

import os
from pathlib import Path

REPO_ROOT = Path(__file__).resolve().parents[2]
UPSTREAM = REPO_ROOT / "third_party" / "PostTrainBench"
ENV_FILE = REPO_ROOT / ".env"


def load_env_file() -> None:
    """Export KEY=VALUE lines from `.env`; variables already set in the shell win."""
    if not ENV_FILE.is_file():
        return
    for line in ENV_FILE.read_text().splitlines():
        key, sep, value = line.strip().partition("=")
        if sep and not key.startswith("#"):
            os.environ.setdefault(key.strip(), value.strip().strip("\"'"))


load_env_file()

DATA_DIR = Path(os.environ.get("PTB_DATA_DIR") or REPO_ROOT / "data").resolve()
CONTAINERS = DATA_DIR / "containers"
HF_CACHE = DATA_DIR / "hf_cache"
# A sweep can opt into a wholly separate result root.  This is stronger than relying only on the
# agent/model prefix below `results/`, and is useful when comparing agents without any possibility
# of touching another agent's artifacts.
RESULTS_DIR = Path(os.environ.get("PTB_RESULTS_DIR") or DATA_DIR / "results").resolve()

# Upstream picks a container per cell with POST_TRAIN_BENCH_CONTAINER_NAME in its `.sub` files:
# `opus_5` for the claude-opus-5 cells, `vllm_debug` for evaluation. Both pin the agent CLIs and
# add the inspect_ai fork that streams vLLM output to the log; plain `standard` does neither.
# `opus_5_local` is upstream's `opus_5` minus two CLI installs whose CDNs the aws cluster cannot
# reach; see the header of containers/opus_5_local.def.
AGENT_CONTAINER = "opus_5_local"
EVAL_CONTAINER = "vllm_debug"
LOCAL_CONTAINERS = REPO_ROOT / "containers"
LOCAL_AGENTS = REPO_ROOT / "agents"


def image(container: str) -> Path:
    return CONTAINERS / f"{container}.sif"


def definition(container: str) -> Path:
    """Local definitions shadow upstream's, so machine-specific variants stay out of the submodule."""
    local = LOCAL_CONTAINERS / f"{container}.def"
    return local if local.is_file() else UPSTREAM / "containers" / f"{container}.def"


def agent_dir(agent: str) -> Path:
    """Return a local agent scaffold when present, otherwise the upstream scaffold."""
    local = LOCAL_AGENTS / agent
    return local if (local / "solve.sh").is_file() else UPSTREAM / "agents" / agent

TASKS = sorted(p.parent.name for p in (UPSTREAM / "src/eval/tasks").glob("*/evaluate.py"))
AGENTS = sorted({p.parent.name for p in (UPSTREAM / "agents").glob("*/solve.sh")}
                | {p.parent.name for p in LOCAL_AGENTS.glob("*/solve.sh")})
DEFINITIONS = sorted({p.stem for p in (UPSTREAM / "containers").glob("*.def")}
                     | {p.stem for p in LOCAL_CONTAINERS.glob("*.def")})


def judge_pool_targets() -> list[str]:
    """`host:port` of every ready replica of the local vLLM judge pool, or an empty list.

    Stands in for OpenAI's `gpt-5-mini` on the two LLM-graded benchmarks, because `.env` has no
    `OPENAI_API_KEY` and api.openai.com is not routable from the aws cluster. Either name the
    replicas with `PTB_JUDGE_POOL`, or point `PTB_JUDGE_POOL_DIR` at the pool directory that
    `scripts/ptb_judge_pool.sbatch` fills with one `task_<i>.addr` per replica -- it writes each
    file only after that endpoint answers, so a reader never sees a half-started server.
    """
    if explicit := os.environ.get("PTB_JUDGE_POOL", "").replace(",", " ").split():
        return explicit
    pool_dir = os.environ.get("PTB_JUDGE_POOL_DIR", "")
    if not pool_dir:
        return []
    return sorted(path.read_text().strip() for path in Path(pool_dir).glob("task_*.addr"))