File size: 17,938 Bytes
565bf0c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 | """
脚本名称: rewrite.py
功能: 批量重写 Task 1, 2, 3 数据集中的 CoT (Chain of Thought) 推理文本。
【功能描述】
该脚本读取原始 JSONL 文件,保持图像路径、几何参数、正确答案不变,
仅根据 metadata 中的几何信息(角度、步骤)重新生成 'cot_trace' 字段。
支持通过修改脚本顶部的模板列表来丰富语言的多样性。
【环境依赖】
python >= 3.6
tqdm
【如何运行】
在终端中使用以下命令运行。请确保 --input_dir 指向包含 jsonl 文件的根目录。
1. 针对 Task 1 (单步预测):
python scripts/rewrite.py --task_type task1 --input_dir ./data/task1 --output_dir ./data/task1_new
2. 针对 Task 2 (多步指令跟随):
python scripts/rewrite.py --task_type task2 --input_dir ./data/task2 --output_dir ./data/task2_new
3. 针对 Task 3 (序列排序):
python scripts/rewrite.py --task_type task3 --input_dir ./data/task3 --output_dir ./data/task3_new
【参数说明】
--input_dir : 输入文件夹路径(脚本会递归查找该目录下的所有 .jsonl 文件)
--output_dir: 输出文件夹路径(保持原有的目录结构)
--task_type : 任务类型,必须是 [task1, task2, task3] 之一
--seed : 随机种子,用于控制模板选择的随机性 (默认 42)
--debug : 开启调试模式,打印错误信息
"""
import os
import json
import re
import argparse
from glob import glob
from tqdm import tqdm
import random
# ==========================================
# 1. 语言模板库 (Language Templates)
# ==========================================
#
# 【如何增加/修改模板】
# 1. 每个列表包含多个字符串,脚本会随机选择其中一条。
# 2. 必须保留大括号 {} 包裹的占位符,例如 {angle}, {direction}。
# 3. 如果你想增加新的表达方式,直接在对应的列表中添加字符串即可。
#
# 【通用变量说明】
# {angle} : 旋转的角度数值 (绝对值)
# {direction}: 旋转方向 (如 clockwise, to the left)
# {img_tag} : 图片标签,格式为 <image_start>[reasoning_image_x]<image_end>
# {label} : 正确选项 (A, B, C, D)
# ==========================================
# --- 通用方向词汇映射 ---
# 用于将 metadata 中的 "clockwise" 替换为更多样的表达
DIRECTION_MAP = {
"clockwise": ["clockwise", "to the right", "in a clockwise direction"],
"anticlockwise": ["anticlockwise", "counter-clockwise", "to the left"],
"counter-clockwise": ["anticlockwise", "counter-clockwise", "to the left"]
}
# ---------------------------------------------------------
# Task 1 模板: 单步预测 (Single-step View Prediction)
# 逻辑:
# 1. Start: 描述初始旋转。
# 2. Middle: 描述中间的连续旋转步骤。
# 3. Final: 总结最终视图并匹配选项。
# ---------------------------------------------------------
T1_START = [
"Starting from the initial view, I rotate the camera {angle} degrees {direction} and see {img_tag}",
"First, let's move the camera {angle} degrees {direction}. The object now looks like this: {img_tag}",
"Initiating a {direction} rotation of {angle} degrees reveals this perspective: {img_tag}",
]
T1_MIDDLE = [
"Continuing the rotation by {angle} degrees {direction}, the view becomes {img_tag}",
"Another {angle} degrees {direction} turn brings us to this angle: {img_tag}",
"Rotating further by {angle} degrees {direction}, I observe {img_tag}",
]
T1_FINAL = [
", which represents the final target view. Comparing this with the options, image {label} is the best match, so the answer is {label}.",
". This matches the final position. Upon checking the candidates, option {label} aligns perfectly with this view. Therefore, the correct answer is {label}.",
", arriving at the destination angle. Among the choices, option {label} is identical to my current view. Thus, {label} is correct."
]
# ---------------------------------------------------------
# Task 2 模板: 多步指令 (Multi-step Instruction Following)
# 逻辑:
# 1. Step (Intermediate): 执行指令 -> 展示 reasoning_image。
# 2. Final Step: 执行最后一步指令 -> 到达目标位置 (注意:最后一步通常没有 reasoning_image,直接对应选项)。
# 3. Conclusion: 匹配选项。
# ---------------------------------------------------------
T2_STEP = [
"Step {i}: Rotating {angle} degrees {direction}, the view transitions to <image_start>[{img_key}]<image_end>",
"Following the instruction to rotate {angle} degrees {direction}, I observe this intermediate view: <image_start>[{img_key}]<image_end>",
"Next, a {angle}-degree {direction} rotation reveals: <image_start>[{img_key}]<image_end>",
]
T2_FINAL_STEP = [
"Finally, rotating {angle} degrees {direction} brings us to the target position.",
"The last step is a {angle}-degree {direction} rotation to reach the destination.",
"Completing the sequence with a {angle} degrees {direction} turn.",
]
T2_CONCLUSION = [
" Comparing the final view with the options, it matches option {label}. So the answer is {label}.",
" This final perspective corresponds to option {label}. Therefore, {label} is correct.",
]
# ---------------------------------------------------------
# Task 3 模板: 序列排序 (View Ordering)
# 逻辑:
# 1. Start: 确定方向,开始旋转。
# 2. Middle (No Match): 旋转后展示图片,但该图片不对应任何选项图片 (只是中间过程)。
# 3. Middle (Match): 旋转后展示图片,并且该图片与选项中的某张图 (1/2/3/4) 匹配。
# 4. Conclusion: 总结正确的顺序 (如 4-2-1-3) 并选择选项。
# ---------------------------------------------------------
T3_START = [
"Based on the images, the rotation appears to be {direction}. Starting the rotation by {angle} degrees, I see <image_start>[{img_key}]<image_end>",
"I deduce the rotation is {direction}. First, moving {angle} degrees reveals <image_start>[{img_key}]<image_end>",
]
# 当这一步的 reasoning_image 不匹配任何选项图片时使用:
T3_MIDDLE_NO_MATCH = [
"Continuing {angle} degrees {direction}, the view is <image_start>[{img_key}]<image_end>",
"Rotating another {angle} degrees {direction} shows <image_start>[{img_key}]<image_end>",
"Next, moving {angle} degrees {direction} gives us <image_start>[{img_key}]<image_end>",
]
# 当这一步的 reasoning_image 匹配了选项图片 (img_idx) 时使用:
# {img_idx} 是匹配到的图片编号 (1, 2, 3, 4)
T3_MIDDLE_MATCH = [
"After rotating {angle} degrees {direction}, I see <image_start>[{img_key}]<image_end>. This view closely resembles image {img_idx}, so the next item in the sequence is {img_idx}.",
"Moving {angle} degrees {direction} leads to <image_start>[{img_key}]<image_end>, which matches image {img_idx}. Thus, {img_idx} is the next step.",
"A further {angle} degrees {direction} rotation shows <image_start>[{img_key}]<image_end>. This looks identical to image {img_idx}.",
]
T3_CONCLUSION = [
"Combining these observations, the correct chronological order is {seq_str}. This corresponds to option {label}.",
"Therefore, the sequence is {seq_str}, making {label} the correct choice.",
"So, the sequence should be {seq_str}, which indicates that option {label} should be the right answer."
]
# ==========================================
# 2. 核心处理逻辑
# ==========================================
def get_direction_variations(direction_str):
"""根据方向关键词返回同义词列表"""
key = direction_str.lower()
if "counter" in key or "anti" in key:
return DIRECTION_MAP["anticlockwise"]
return DIRECTION_MAP["clockwise"]
def process_task1(line_data):
"""处理 Task 1: 单步预测"""
metadata = line_data.get("metadata", {})
old_cot = metadata.get("cot_trace", "")
direction = metadata.get("direction", "clockwise")
gt_answer = line_data.get("gt_answer", "")
label_match = re.search(r'<answer>([A-D])</answer>', gt_answer)
correct_label = label_match.group(1) if label_match else "A"
# 从旧文本中提取所有数字作为角度步骤
steps = [int(m) for m in re.findall(r'(\d+)\s+degrees', old_cot)]
images = line_data.get("images", {})
# 按索引排序 reasoning images
reasoning_keys = sorted([k for k in images.keys() if k.startswith("reasoning_image_")],
key=lambda x: int(x.split('_')[-1]))
# 简单校验:步骤数应等于中间图数量
if len(steps) != len(reasoning_keys):
return line_data, False
cot_parts = []
dir_vars = get_direction_variations(direction)
for i, (step, img_key) in enumerate(zip(steps, reasoning_keys)):
img_tag = f"<image_start>[{img_key}]<image_end>"
cur_dir = random.choice(dir_vars)
if i == 0:
tmpl = random.choice(T1_START)
else:
tmpl = random.choice(T1_MIDDLE)
cot_parts.append(tmpl.format(angle=step, direction=cur_dir, img_tag=img_tag))
new_cot = "; ".join(cot_parts)
new_cot += random.choice(T1_FINAL).format(label=correct_label)
line_data['metadata']['cot_trace'] = new_cot
return line_data, True
def process_task2(line_data):
"""处理 Task 2: 多步指令"""
metadata = line_data.get("metadata", {})
# 优先从 metadata 获取准确的角度列表
steps_degrees = metadata.get("steps_degrees", [])
if not steps_degrees:
instr = metadata.get("instruction_sequence", "")
steps_degrees = [int(m) for m in re.findall(r'(\d+)\s+degrees', instr)]
instruction_seq = metadata.get("instruction_sequence", "")
# 提取指令中的方向序列
directions = re.findall(r'(clockwise|anticlockwise|counter-clockwise)', instruction_seq)
gt_answer = line_data.get("gt_answer", "")
label_match = re.search(r'<answer>([A-D])</answer>', gt_answer)
correct_label = label_match.group(1) if label_match else "D"
images = line_data.get("images", {})
reasoning_keys = sorted([k for k in images.keys() if k.startswith("reasoning_image_")],
key=lambda x: int(x.split('_')[-1]))
# Task 2 逻辑:N 个步骤,通常有 N-1 张中间图 (最后一步直接到结果)
if len(steps_degrees) != len(reasoning_keys) + 1:
# 容错:如果步骤数不匹配,不修改
return line_data, False
cot_parts = []
for i, angle in enumerate(steps_degrees):
abs_angle = abs(angle)
# 获取当前步骤对应的方向
cur_dir_raw = directions[i] if i < len(directions) else "clockwise"
cur_dir = random.choice(get_direction_variations(cur_dir_raw))
if i < len(reasoning_keys):
# 中间步骤:有 reasoning_image
img_key = reasoning_keys[i]
tmpl = random.choice(T2_STEP)
cot_parts.append(tmpl.format(i=i+1, angle=abs_angle, direction=cur_dir, img_key=img_key))
else:
# 最后一步:没有 reasoning_image,直接得出结论
tmpl = random.choice(T2_FINAL_STEP)
cot_parts.append(tmpl.format(angle=abs_angle, direction=cur_dir))
new_cot = " ".join(cot_parts)
new_cot += random.choice(T2_CONCLUSION).format(label=correct_label)
line_data['metadata']['cot_trace'] = new_cot
return line_data, True
def process_task3(line_data):
"""处理 Task 3: 序列排序"""
metadata = line_data.get("metadata", {})
old_cot = metadata.get("cot_trace", "")
# 1. 确定总体旋转方向
if "counter-clockwise" in old_cot or "anticlockwise" in old_cot:
direction_raw = "anticlockwise"
else:
direction_raw = "clockwise"
gt_answer = line_data.get("gt_answer", "")
label_match = re.search(r'<answer>([A-D])</answer>', gt_answer)
correct_label = label_match.group(1) if label_match else "A"
# 2. 提取最终的排序结果 (如 "4-2-1-3")
seq_match = re.search(r'sequence should be ([\d-]+)', old_cot)
final_seq_str = seq_match.group(1) if seq_match else "UNKNOWN"
# 3. 解析旧 CoT,提取 (角度, 图片key, 是否匹配) 的三元组
# 策略:按 "Then," 或 "After" 分割句子,逐句分析
segments = re.split(r'(?:Then,|After)', old_cot)
parsed_steps = []
for seg in segments:
# 提取角度
angle_m = re.search(r'rotat\w+\s+(\d+)\s+degrees', seg)
if not angle_m: continue
angle = int(angle_m.group(1))
# 提取 reasoning_image 编号
img_m = re.search(r'reasoning_image_(\d+)', seg)
if not img_m: continue
r_img_idx = img_m.group(1)
r_img_key = f"reasoning_image_{r_img_idx}"
# 提取匹配信息 (matches image X)
match_m = re.search(r'(?:matches|resembles)\s+image\s+(\d+)', seg)
matched_idx = match_m.group(1) if match_m else None
parsed_steps.append({
"angle": angle,
"img_key": r_img_key,
"matched_idx": matched_idx
})
images = line_data.get("images", {})
reasoning_keys = [k for k in images.keys() if k.startswith("reasoning_image_")]
# 校验解析出的步骤数是否与图片数一致
if len(parsed_steps) != len(reasoning_keys):
return line_data, False
# 4. 生成新文本
cot_parts = []
dir_vars = get_direction_variations(direction_raw)
for i, step in enumerate(parsed_steps):
cur_dir = random.choice(dir_vars)
angle = step['angle']
img_key = step['img_key']
matched_idx = step['matched_idx']
if i == 0:
# 第一步通常只是展示,不匹配
tmpl = random.choice(T3_START)
text = tmpl.format(angle=angle, direction=cur_dir, img_key=img_key)
else:
if matched_idx:
# 如果这一步有匹配
tmpl = random.choice(T3_MIDDLE_MATCH)
text = tmpl.format(angle=angle, direction=cur_dir, img_key=img_key, img_idx=matched_idx)
else:
# 如果这一步只是中间过渡
tmpl = random.choice(T3_MIDDLE_NO_MATCH)
text = tmpl.format(angle=angle, direction=cur_dir, img_key=img_key)
cot_parts.append(text)
new_cot = " ".join(cot_parts)
new_cot += random.choice(T3_CONCLUSION).format(seq_str=final_seq_str, label=correct_label)
line_data['metadata']['cot_trace'] = new_cot
return line_data, True
# ==========================================
# 3. 主程序入口
# ==========================================
def process_file(file_path, out_path, task_type, debug=False):
new_lines = []
modified_count = 0
total_count = 0
with open(file_path, 'r', encoding='utf-8') as f:
for line in f:
if not line.strip(): continue
total_count += 1
data = json.loads(line)
is_modified = False
try:
if task_type == "task1":
data, is_modified = process_task1(data)
elif task_type == "task2":
data, is_modified = process_task2(data)
elif task_type == "task3":
data, is_modified = process_task3(data)
except Exception as e:
if debug: print(f"[Error] Line {total_count} in {os.path.basename(file_path)}: {e}")
is_modified = False
if is_modified:
modified_count += 1
new_lines.append(data)
with open(out_path, 'w', encoding='utf-8') as f_out:
for item in new_lines:
f_out.write(json.dumps(item) + "\n")
return total_count, modified_count
def main():
parser = argparse.ArgumentParser(description="Rewrite CoT trace for Task 1, 2, 3")
parser.add_argument("--input_dir", type=str, required=True, help="Input directory root (recursive search)")
parser.add_argument("--output_dir", type=str, required=True, help="Output directory root")
parser.add_argument("--task_type", type=str, required=True, choices=["task1", "task2", "task3"],
help="Which task logic to apply")
parser.add_argument("--seed", type=int, default=42, help="Random seed for template selection")
parser.add_argument("--debug", action="store_true", help="Print detailed error messages")
args = parser.parse_args()
random.seed(args.seed)
# 递归查找所有 jsonl 文件
search_pattern = os.path.join(args.input_dir, "**", "*.jsonl")
files = glob(search_pattern, recursive=True)
if not files:
print(f"No JSONL files found in {args.input_dir}")
return
print(f"Found {len(files)} files for {args.task_type}")
print(f"Input: {args.input_dir}")
print(f"Output: {args.output_dir}")
total_processed = 0
total_modified = 0
for file_path in tqdm(files, desc=f"Processing {args.task_type}"):
# 计算相对路径,保持输出目录结构一致
rel_path = os.path.relpath(file_path, args.input_dir)
out_path = os.path.join(args.output_dir, rel_path)
os.makedirs(os.path.dirname(out_path), exist_ok=True)
t, m = process_file(file_path, out_path, args.task_type, args.debug)
total_processed += t
total_modified += m
print("-" * 30)
print(f"Done!")
print(f"Total Lines Processed: {total_processed}")
print(f"Total Lines Modified : {total_modified}")
print(f"Success Rate: {total_modified/total_processed:.1%}" if total_processed > 0 else "N/A")
print(f"Output saved to {args.output_dir}")
if __name__ == "__main__":
main()
|