File size: 5,914 Bytes
6aeb377 1ce72ab 6aeb377 1ce72ab 6aeb377 1ce72ab 6aeb377 1ce72ab 6aeb377 1ce72ab 6aeb377 1ce72ab 6aeb377 1ce72ab 6aeb377 | 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 | from __future__ import annotations
import argparse
import json
import shutil
import tempfile
from pathlib import Path
from huggingface_hub import HfApi
METADATA_DEFAULTS = {
"agent_id": "",
"arena_version": "",
"dataset_version": "",
"generator_version": "",
"label_source": "",
"phase": "",
"prompt_version": "",
"seed": 0,
"split": "",
"policy_reward": 0.0,
"solver_joint_actions_explored": 0,
"solver_optimal_count": 0,
"solver_reward": 0.0,
"generator_mode": "",
"targeted_skill": "",
}
DATASET_CARD = """---
license: apache-2.0
task_categories:
- text-generation
language:
- en
configs:
- config_name: default
data_files:
- split: train_broadcast
path: train_broadcast.jsonl
- split: train_action
path: train_action_common.jsonl
- split: train_action_rare
path: train_action_rare.jsonl
- split: validation
path: validation.jsonl
- split: test
path: test.jsonl
- split: overfit
path: overfit.jsonl
---
# Swarm Arena SFT v2
Solver-filtered warm-start data for the deterministic Swarm Arena 4v4
coordination environment. Each row contains system, user, and assistant messages
plus provenance metadata. Training broadcasts and actions are separate splits so
sampling can preserve a 60/40 phase mixture. Validation and test are never
reweighted.
The simulator, oracle, audit, frozen evaluation, and Prime-RL configs live in
<https://github.com/ChinmayK0607/blog-rl/tree/exp/swarm-arena-4b/experiments/swarm_arena>.
Optional provenance fields use typed zero/empty-string defaults so every JSONL
split has one stable Arrow schema. They do not affect prompts or targets.
Dataset content SHA-256:
`edad09bb301748621a0fab73ebf3de60d60abfd9f56c9afcc6ca02ffe12f3a80`.
"""
def normalize_row(row: dict) -> dict:
row = dict(row)
row["metadata"] = {**METADATA_DEFAULTS, **row["metadata"]}
return row
def write_normalized_split(source: Path, destination: Path) -> None:
with source.open(encoding="utf-8") as input_handle, destination.open(
"w", encoding="utf-8"
) as output_handle:
for line in input_handle:
output_handle.write(json.dumps(normalize_row(json.loads(line)), sort_keys=True) + "\n")
def split_train(
source: Path,
broadcast_path: Path,
common_action_path: Path,
rare_action_path: Path,
) -> None:
rare_types = {"WAIT", "SCAN", "TRANSFER"}
with source.open(encoding="utf-8") as input_handle, broadcast_path.open(
"w", encoding="utf-8"
) as broadcast_handle, common_action_path.open(
"w", encoding="utf-8"
) as common_handle, rare_action_path.open("w", encoding="utf-8") as rare_handle:
for line in input_handle:
row = normalize_row(json.loads(line))
if row["metadata"]["phase"] == "BROADCAST":
destination = broadcast_handle
else:
user = json.loads(row["messages"][-2]["content"])
action_id = json.loads(row["messages"][-1]["content"])["action_id"]
action = next(item for item in user["legal_actions"] if item["id"] == action_id)
destination = rare_handle if action["type"] in rare_types else common_handle
destination.write(json.dumps(row, sort_keys=True) + "\n")
def make_overfit(source: Path, destination: Path, count: int = 256) -> None:
rows = [json.loads(line) for line in source.read_text(encoding="utf-8").splitlines()]
broadcasts = [row for row in rows if row["metadata"]["phase"] == "BROADCAST"][:154]
actions = [row for row in rows if row["metadata"]["phase"] == "ACT"]
rare_types = {"WAIT", "SCAN", "TRANSFER"}
def is_rare(row: dict) -> bool:
user = json.loads(row["messages"][-2]["content"])
action_id = json.loads(row["messages"][-1]["content"])["action_id"]
action = next(item for item in user["legal_actions"] if item["id"] == action_id)
return action["type"] in rare_types
rare = [row for row in actions if is_rare(row)][:24]
common = [row for row in actions if not is_rare(row)][: count - 154 - len(rare)]
with destination.open("w", encoding="utf-8") as handle:
for row in broadcasts + rare + common:
handle.write(json.dumps(normalize_row(row), sort_keys=True) + "\n")
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("source", type=Path)
parser.add_argument("--repo-id", default="CK0607/swarm-arena-sft-v2")
parser.add_argument("--experiment-root", type=Path)
args = parser.parse_args()
api = HfApi()
api.create_repo(args.repo_id, repo_type="dataset", exist_ok=True)
with tempfile.TemporaryDirectory(prefix="swarm-arena-hf-") as temporary:
staging = Path(temporary)
split_train(
args.source / "train.jsonl",
staging / "train_broadcast.jsonl",
staging / "train_action_common.jsonl",
staging / "train_action_rare.jsonl",
)
make_overfit(args.source / "train.jsonl", staging / "overfit.jsonl")
for filename in ("validation.jsonl", "test.jsonl"):
write_normalized_split(args.source / filename, staging / filename)
for filename in ("manifest.json", "audit.json"):
shutil.copy2(args.source / filename, staging / filename)
(staging / "README.md").write_text(DATASET_CARD, encoding="utf-8")
if args.experiment_root:
shutil.copytree(
args.experiment_root,
staging / "code",
ignore=shutil.ignore_patterns("__pycache__", "*.pyc", "data", ".venv", "uv.lock"),
)
api.upload_folder(
repo_id=args.repo_id,
repo_type="dataset",
folder_path=staging,
commit_message="Publish audited Swarm Arena SFT v2",
)
if __name__ == "__main__":
main()
|