File size: 9,551 Bytes
9589849 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 | #!/usr/bin/env python3
"""Convert public mini-swe-agent trajectories to Pi/Qwen structured tool chats."""
from __future__ import annotations
import argparse
import json
import re
from pathlib import Path
from datasets import Dataset, load_dataset
SOURCE = "Kwai-Klear/SWE-smith-mini_swe_agent_plus-trajectories-66k"
BASH_BLOCK = re.compile(r"```(?:bash|sh)\s*\n(.*?)\n```", re.DOTALL | re.IGNORECASE)
FINAL_MARKER = "MINI_SWE_AGENT_FINAL_OUTPUT"
PI_SYSTEM_PROMPT = """You are an expert coding assistant operating inside pi, a coding agent harness. You help users by reading files, executing commands, editing code, and writing new files.
Available tools:
- read: Read file contents
- bash: Execute bash commands
- edit: Make precise file edits with exact text replacement
- write: Create or overwrite files
Guidelines:
- Inspect the repository and relevant instructions before changing code.
- Use tools to make the requested change; do not merely describe commands.
- Prefer focused searches and reads over dumping large files.
- Make minimal, targeted edits and preserve unrelated behavior.
- Run relevant focused tests after editing.
- When the task is complete, stop using tools and give a concise final summary.
- Be concise in your responses.
- Show file paths clearly when working with files."""
PI_TOOLS = [
{
"type": "function",
"function": {
"name": "read",
"description": "Read the contents of a file.",
"parameters": {
"type": "object",
"properties": {
"path": {"type": "string"},
"offset": {"type": "number"},
"limit": {"type": "number"},
},
"required": ["path"],
},
},
},
{
"type": "function",
"function": {
"name": "bash",
"description": "Execute a bash command in the current working directory.",
"parameters": {
"type": "object",
"properties": {
"command": {"type": "string"},
"timeout": {"type": "number"},
},
"required": ["command"],
},
},
},
{
"type": "function",
"function": {
"name": "edit",
"description": "Edit one file using exact text replacement.",
"parameters": {
"type": "object",
"properties": {
"path": {"type": "string"},
"edits": {
"type": "array",
"items": {
"type": "object",
"properties": {
"oldText": {"type": "string"},
"newText": {"type": "string"},
},
"required": ["oldText", "newText"],
},
},
},
"required": ["path", "edits"],
},
},
},
{
"type": "function",
"function": {
"name": "write",
"description": "Create or overwrite a file.",
"parameters": {
"type": "object",
"properties": {
"path": {"type": "string"},
"content": {"type": "string"},
},
"required": ["path", "content"],
},
},
},
]
TOOLS_JSON = json.dumps(PI_TOOLS, separators=(",", ":"))
def message(
role: str,
content: str,
*,
reasoning_content: str | None = None,
tool_calls: list[dict] | None = None,
tool_call_id: str | None = None,
name: str | None = None,
) -> dict:
return {
"role": role,
"content": content,
"reasoning_content": reasoning_content,
"tool_calls": tool_calls,
"tool_call_id": tool_call_id,
"name": name,
}
def clean_reasoning(text: str) -> str:
text = text.strip()
text = re.sub(r"^THOUGHT:\s*", "", text, flags=re.IGNORECASE)
return text.strip()
def convert(example: dict) -> dict:
source_messages = example["messages"]
converted = [message("system", PI_SYSTEM_PROMPT)]
pending_call_id: str | None = None
call_index = 0
initial_user_seen = False
malformed_actions = 0
for source in source_messages:
role = source["role"]
content = source.get("content") or ""
if role == "system":
continue
if role == "user":
if pending_call_id is not None:
converted.append(
message(
"tool",
content,
tool_call_id=pending_call_id,
name="bash",
)
)
pending_call_id = None
else:
converted.append(message("user", content))
initial_user_seen = True
continue
if role != "assistant":
continue
match = BASH_BLOCK.search(content)
is_final = FINAL_MARKER in content
if is_final:
converted.append(
message(
"assistant",
"Implemented the requested change and verified it with focused checks.",
)
)
pending_call_id = None
continue
if match is None:
malformed_actions += 1
converted.append(message("assistant", content.strip()))
pending_call_id = None
continue
reasoning = clean_reasoning(content[: match.start()])
command = match.group(1).strip()
call_index += 1
pending_call_id = f"call_{call_index:04d}"
converted.append(
message(
"assistant",
"",
reasoning_content=reasoning,
tool_calls=[
{
"id": pending_call_id,
"type": "function",
"function": {
"name": "bash",
"arguments": json.dumps(
{"command": command}, separators=(",", ":")
),
},
}
],
)
)
if not initial_user_seen:
raise ValueError(f"trajectory {example['instance_id']} has no user request")
if converted[-1]["role"] != "assistant" or converted[-1]["tool_calls"]:
converted.append(
message(
"assistant",
"Implemented the requested change and verified it with focused checks.",
)
)
return {
"instance_id": example["instance_id"],
"messages": converted,
"tools": TOOLS_JSON,
"source": SOURCE,
"num_tool_calls": call_index,
"malformed_actions": malformed_actions,
}
def evaluation_instance_ids(path: Path) -> set[str]:
"""Exact identifiers only, used solely as a decontamination blacklist."""
if not path.exists():
return set()
return {entry.name for entry in path.iterdir() if entry.is_dir()}
def write_shards(dataset: Dataset, output_dir: Path, shards: int) -> None:
output_dir.mkdir(parents=True, exist_ok=True)
for index in range(shards):
shard = dataset.shard(num_shards=shards, index=index, contiguous=True)
path = output_dir / f"train-{index:05d}-of-{shards:05d}.parquet"
shard.to_parquet(path)
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--limit", type=int)
parser.add_argument("--num-proc", type=int, default=32)
parser.add_argument("--shards", type=int, default=16)
parser.add_argument(
"--eval-id-dir",
type=Path,
default=Path("/root/work/shared/tasksets/swe-bench-verified"),
)
args = parser.parse_args()
split = "train" if args.limit is None else f"train[:{args.limit}]"
raw = load_dataset(SOURCE, split=split)
blacklist = evaluation_instance_ids(args.eval_id_dir)
before = len(raw)
raw = raw.filter(
lambda instance_id: instance_id not in blacklist,
input_columns=["instance_id"],
num_proc=args.num_proc,
desc="Removing exact evaluation identifiers",
)
removed = before - len(raw)
converted = raw.map(
convert,
remove_columns=raw.column_names,
num_proc=args.num_proc,
desc="Converting trajectories to Pi structured calls",
)
write_shards(converted, args.output, min(args.shards, len(converted)))
report = {
"source": SOURCE,
"source_revision": "750b2c11239fd5e32f97e6cfb9bf80fb9a9a2983",
"source_rows": before,
"rows_removed_as_exact_eval_ids": removed,
"output_rows": len(converted),
"tool_calls": sum(converted["num_tool_calls"]),
"malformed_actions": sum(converted["malformed_actions"]),
"transformation": "fenced bash -> Pi structured bash calls; terminal marker -> concise final response",
}
(args.output / "provenance.json").write_text(json.dumps(report, indent=2) + "\n")
print(json.dumps(report, indent=2))
if __name__ == "__main__":
main()
|