Files
hack-house/bench/tasks.py
T
leetcrypt 58d405c518 test(ai): baseline 4 local CPU models + fix prompt-fairness bug
Capture native-harness benchmark baselines for qwen2.5 0.5b/1.5b/3b and
qwen2.5-coder:7b (all probed tool-capable; deepseek-r1 and NL2SH reject the
tools field). All cluster at 1-2/12 with high variance; the 7B buys no
pass-rate gain at ~3x latency, so qwen2.5:3b stays the default. The single
biggest score sink across every model is bare tool-call-as-text leaks — a
harness parse gap, the clear next improvement.

Also drop "in your home directory" from the shell prompts: it made literal-
minded models create a home/ subdir (/root/home/a/b/c/...) or use ~/, which
the benchmark itself surfaced. Findings doc + bench README carry the model
comparison table and recommendations (llama3.2:3b / llama3.1:8b for a non-qwen
data point).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-06-10 09:30:36 -07:00

163 lines
6.9 KiB
Python

"""Native-harness benchmark task matrix.
Each task exercises the `/ai <model> !<task>` native tool-calling loop
(cmd_chat/agent/bridge.py:_run_native) end-to-end and is graded by GROUND
TRUTH inside the sandbox container — never by the model's own summary, which
the weak CPU model is known to fabricate (see
docs/findings-native-harness-2026-06-10.md).
Four categories, each with an easy / medium / hard tier:
shell — run_shell only (filesystem side effects)
code — write_file + run_shell (author a program, execute it, capture output)
git — clone/inspect a remote repo (needs container network + git)
multi — multi-step / recovery chains (the loop's hardest cases)
Conventions every task obeys so grading is deterministic:
* Tasks operate in the agent's real cwd — the container HOME (/root) — and
prompts use BARE filenames ("write who.txt"). The weak model writes
`run_shell` redirects relative to its cwd, so pinning an absolute
/root/bench path instead just measures absolute-path discipline (which it
fails ~always) and drowns out every other signal. Verify snippets use the
absolute /root/<name> form so grading is unambiguous.
* `CLEAN` wipes ONLY the suite's known artifacts before each task — it never
`rm -rf`s the home directory.
* `verify` is a POSIX-sh snippet run via `podman exec`; exit 0 == PASS.
* `setup` (optional) is a per-task sh snippet run before the task (e.g. to
seed an input file the task is supposed to consume).
"""
from dataclasses import dataclass, field
HOME = "/root"
# Every file/dir the suite can create. Only these are wiped before each task,
# so one task can't leak state into the next without nuking the home dir.
ARTIFACTS = [
"who.txt", "a", "conf_count.txt", "add.py", "add.out", "greet.sh",
"greet.out", "fib.py", "fib.out", "hw", "hw2", "hw3", "readme.txt",
"branch.txt", "proj", "fruit.txt", "sorted.txt", "words.txt", "wc.py",
"wc.out",
]
CLEAN = f"cd {HOME} && rm -rf " + " ".join(ARTIFACTS)
@dataclass(frozen=True)
class Task:
id: str
category: str # shell | code | git | multi
tier: str # easy | medium | hard
prompt: str # the text sent after `/ai <model> !`
verify: str # sh snippet; exit 0 == task succeeded (ground truth)
setup: str = "" # sh snippet run before the task (seed inputs)
needs_net: bool = field(default=False)
# fmt: off
TASKS = [
# ───────────────────────── shell ─────────────────────────
Task(
"shell-easy", "shell", "easy",
"run whoami and write its output to a file named who.txt",
f'test "$(cat {HOME}/who.txt)" = root',
),
Task(
"shell-medium", "shell", "medium",
"create the nested directory a/b/c and write the text OK into "
"a/b/c/marker.txt",
f'test "$(cat {HOME}/a/b/c/marker.txt)" = OK',
),
Task(
"shell-hard", "shell", "hard",
"count how many files under /etc end in .conf and write just that number "
"to a file conf_count.txt",
f'test "$(cat {HOME}/conf_count.txt)" = "$(find /etc -name "*.conf" -type f 2>/dev/null | wc -l | tr -d " ")"',
),
# ───────────────────────── code ──────────────────────────
Task(
"code-easy", "code", "easy",
"write a Python file add.py that prints the result of 2+3, run it with "
"python3 and save the output to add.out",
f'test "$(cat {HOME}/add.out)" = 5',
),
Task(
# the documented give-up case: write + chmod +x + ./run. A weak model
# tends to skip chmod, hit exit 126, and quit — the nudge loop's target.
"code-medium", "code", "medium",
"write a shell script greet.sh that echoes hello, make it executable with "
"chmod, run it as ./greet.sh and save the output to greet.out",
f'test -x {HOME}/greet.sh && test "$(cat {HOME}/greet.out)" = hello',
),
Task(
"code-hard", "code", "hard",
"write a Python script fib.py that prints the first 10 Fibonacci numbers "
"(starting 0 1) space-separated on one line, run it and save the output "
"to fib.out",
f'test "$(cat {HOME}/fib.out)" = "0 1 1 2 3 5 8 13 21 34"',
),
# ───────────────────────── git ───────────────────────────
Task(
"git-easy", "git", "easy",
"git clone https://github.com/octocat/Hello-World into a directory hw",
f'test -d {HOME}/hw/.git',
needs_net=True,
),
Task(
"git-medium", "git", "medium",
"clone https://github.com/octocat/Hello-World into hw2, then write the "
"contents of its README to readme.txt",
f'grep -qi "hello world" {HOME}/readme.txt',
needs_net=True,
),
Task(
"git-hard", "git", "hard",
"clone https://github.com/octocat/Hello-World into hw3, then write the "
"name of its current branch to branch.txt",
f'test "$(tr -d " \n" < {HOME}/branch.txt)" = master',
needs_net=True,
),
# ──────────────────────── multi ──────────────────────────
Task(
# validated multi-step case (mkdir -> write -> list) that used to stall
"multi-easy", "multi", "easy",
"make a directory proj, create proj/notes.txt containing the text hello "
"world, then list proj",
f'test "$(cat {HOME}/proj/notes.txt)" = "hello world"',
),
Task(
"multi-medium", "multi", "medium",
"the file fruit.txt has the lines apple, banana, cherry; write those "
"lines reverse-sorted into sorted.txt",
setup=f'printf "apple\\nbanana\\ncherry\\n" > {HOME}/fruit.txt',
verify=f'test "$(cat {HOME}/sorted.txt)" = "$(printf \'cherry\\nbanana\\napple\')"',
),
Task(
"multi-hard", "multi", "hard",
"the file words.txt contains 'one two three'; write a Python script wc.py "
"that counts the words in that file and prints only the number, run it "
"and save the output to wc.out",
setup=f'printf "one two three\\n" > {HOME}/words.txt',
verify=f'test "$(tr -d " \n" < {HOME}/wc.out)" = 3',
),
]
# fmt: on
def select(ids=None, categories=None, tiers=None, allow_net=True):
"""Filter the matrix by id / category / tier, optionally dropping net tasks."""
out = []
for t in TASKS:
if ids and t.id not in ids:
continue
if categories and t.category not in categories:
continue
if tiers and t.tier not in tiers:
continue
if t.needs_net and not allow_net:
continue
out.append(t)
return out