File size: 9,032 Bytes
74ab802 | 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 | #!/usr/bin/env python3
"""
Extract divergent steps for aligned EB-ALFRED mixed-task cases.
For each aligned success/failure case pair, the script finds the most similar
success run and failure run by maximizing their shared action prefix length.
It then saves only the steps from the first divergence onward for both sides.
"""
from __future__ import annotations
import argparse
import copy
import json
from pathlib import Path
from typing import Any
DEFAULT_SUCCESS_INPUT = Path(
"/data/Top_Spcae/ICML_2026/Dataset/eb-alfred_mixed_tasks_success_cases.json"
)
DEFAULT_FAILURE_INPUT = Path(
"/data/Top_Spcae/ICML_2026/Dataset/eb-alfred_mixed_tasks_failure_cases.json"
)
DEFAULT_OUTPUT = Path(
"/data/Top_Spcae/ICML_2026/Dataset/eb-alfred_mixed_tasks_diff_steps.json"
)
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(
description="Extract divergent steps from aligned EB-ALFRED mixed-task cases."
)
parser.add_argument(
"--success-input",
type=Path,
default=DEFAULT_SUCCESS_INPUT,
help=f"Path to aligned success cases JSON. Default: {DEFAULT_SUCCESS_INPUT}",
)
parser.add_argument(
"--failure-input",
type=Path,
default=DEFAULT_FAILURE_INPUT,
help=f"Path to aligned failure cases JSON. Default: {DEFAULT_FAILURE_INPUT}",
)
parser.add_argument(
"--output",
type=Path,
default=DEFAULT_OUTPUT,
help=f"Path to output JSON. Default: {DEFAULT_OUTPUT}",
)
return parser.parse_args()
def action_sequence(run: dict[str, Any]) -> list[str]:
actions: list[str] = []
for step in run.get("trajectory", []) or []:
executable_plan = step.get("executable_plan", []) or []
if executable_plan:
actions.append(str(executable_plan[0]["action"][1]))
return actions
def shared_prefix_len(a: list[str], b: list[str]) -> int:
prefix = 0
for left, right in zip(a, b):
if left != right:
break
prefix += 1
return prefix
def choose_best_pair(
success_case: dict[str, Any], failure_case: dict[str, Any]
) -> tuple[dict[str, Any], dict[str, Any], int]:
best_success_run: dict[str, Any] | None = None
best_failure_run: dict[str, Any] | None = None
best_prefix = -1
best_success_len = -1
best_failure_len = -1
for success_run in success_case.get("runs", []):
success_actions = action_sequence(success_run)
for failure_run in failure_case.get("runs", []):
failure_actions = action_sequence(failure_run)
prefix = shared_prefix_len(success_actions, failure_actions)
# Prefer longer shared prefixes. On ties, prefer shorter total tails.
success_len = len(success_actions)
failure_len = len(failure_actions)
is_better = (
prefix > best_prefix
or (
prefix == best_prefix
and success_len + failure_len < best_success_len + best_failure_len
)
)
if is_better:
best_success_run = success_run
best_failure_run = failure_run
best_prefix = prefix
best_success_len = success_len
best_failure_len = failure_len
if best_success_run is None or best_failure_run is None:
raise ValueError(
f"Could not find a valid run pair for case {success_case.get('case_index')}"
)
return best_success_run, best_failure_run, best_prefix
def annotate_diff_steps(
trajectory_steps: list[dict[str, Any]],
prefix: int,
trajectory_outcome: str,
selected_model: str | None,
) -> list[dict[str, Any]]:
annotated_steps: list[dict[str, Any]] = []
trajectory_success = 1.0 if trajectory_outcome == "success" else 0.0
for original_step_index, step in enumerate(trajectory_steps[prefix:], start=prefix):
annotated_step = copy.deepcopy(step)
executable_plan = annotated_step.get("executable_plan", []) or []
for executable in executable_plan:
executable["trajectory_outcome"] = trajectory_outcome
executable["trajectory_success"] = trajectory_success
executable["selected_model"] = selected_model
executable["original_step_index_0based"] = original_step_index
executable["original_step_index_1based"] = original_step_index + 1
annotated_step["trajectory_outcome"] = trajectory_outcome
annotated_step["trajectory_success"] = trajectory_success
annotated_step["selected_model"] = selected_model
annotated_step["original_step_index_0based"] = original_step_index
annotated_step["original_step_index_1based"] = original_step_index + 1
annotated_steps.append(annotated_step)
return annotated_steps
def build_diff_case(
success_case: dict[str, Any], failure_case: dict[str, Any]
) -> dict[str, Any]:
keys_to_match = ["case_index", "trajectory_id", "eval_set", "episode_id", "instruction"]
for key in keys_to_match:
if success_case.get(key) != failure_case.get(key):
raise ValueError(
f"Mismatched aligned cases for key '{key}': "
f"{success_case.get(key)!r} != {failure_case.get(key)!r}"
)
success_run, failure_run, prefix = choose_best_pair(success_case, failure_case)
success_actions = action_sequence(success_run)
failure_actions = action_sequence(failure_run)
success_diff_steps = annotate_diff_steps(
success_run.get("trajectory", []) or [],
prefix,
"success",
success_run.get("model_name"),
)
failure_diff_steps = annotate_diff_steps(
failure_run.get("trajectory", []) or [],
prefix,
"failure",
failure_run.get("model_name"),
)
success_first_diff_action = success_actions[prefix] if prefix < len(success_actions) else None
failure_first_diff_action = failure_actions[prefix] if prefix < len(failure_actions) else None
return {
"case_index": success_case["case_index"],
"trajectory_id": success_case["trajectory_id"],
"eval_set": success_case["eval_set"],
"episode_id": success_case["episode_id"],
"instruction": success_case["instruction"],
"paired_success_run_count": success_case["paired_success_run_count"],
"paired_failure_run_count": success_case["paired_failure_run_count"],
"selected_success_model": success_run.get("model_name"),
"selected_failure_model": failure_run.get("model_name"),
"selected_success_input": success_run.get("input"),
"selected_failure_input": failure_run.get("input"),
"selected_success_num_steps": success_run.get("num_steps"),
"selected_failure_num_steps": failure_run.get("num_steps"),
"shared_prefix_len": prefix,
"shared_prefix_actions": success_actions[:prefix],
"first_diff_step_index_0based": prefix,
"first_diff_step_index_1based": prefix + 1,
"first_diff_success_action": success_first_diff_action,
"first_diff_failure_action": failure_first_diff_action,
"success_diff_num_steps": len(success_diff_steps),
"failure_diff_num_steps": len(failure_diff_steps),
"success_diff_actions": success_actions[prefix:],
"failure_diff_actions": failure_actions[prefix:],
"success_diff_steps": success_diff_steps,
"failure_diff_steps": failure_diff_steps,
}
def main() -> None:
args = parse_args()
with args.success_input.open("r", encoding="utf-8") as f:
success_payload = json.load(f)
with args.failure_input.open("r", encoding="utf-8") as f:
failure_payload = json.load(f)
success_cases = success_payload["cases"]
failure_cases = failure_payload["cases"]
if len(success_cases) != len(failure_cases):
raise ValueError(
f"Case count mismatch: {len(success_cases)} success cases vs "
f"{len(failure_cases)} failure cases"
)
diff_cases = [
build_diff_case(success_case, failure_case)
for success_case, failure_case in zip(success_cases, failure_cases)
]
output_payload = {
"success_input_file": str(args.success_input),
"failure_input_file": str(args.failure_input),
"selection_rule": (
"For each aligned case, choose the success/failure run pair with the "
"longest shared action prefix; extract only the divergent tail steps."
),
"num_cases": len(diff_cases),
"cases": diff_cases,
}
args.output.parent.mkdir(parents=True, exist_ok=True)
with args.output.open("w", encoding="utf-8") as f:
json.dump(output_payload, f, ensure_ascii=False, indent=2)
print(f"Aligned cases processed: {len(diff_cases)}")
print(f"Saved divergent-step file to: {args.output}")
if __name__ == "__main__":
main()
|