File size: 2,774 Bytes
1c28b5d | 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 | """Flatten v7 nested-polygon annotations without exposing raw generation metadata."""
import csv,json
from pathlib import Path
TASKS={1:"Image Description",2:"Basic Relational Reasoning",3:"Comparative Reasoning",4:"Compound Reasoning",5:"Extrapolative/Counterfactual Reasoning"}
META_KEYS=("difficulty_score","rotation_mode","reduction_mode","step_reduction_factors","reduction_factor_span","factor_progression_direction","extrapolation_reduction_factor","line_width_px","seed","parameter_seed","symmetry_modulus_degrees","cumulative_rotation_degrees","cumulative_rotation_fraction","target_cumulative_rotation_degrees","generation_attempt","generation_rejections")
def flatten(annotations):
annotations=Path(annotations);root=annotations.parent
records=[json.loads(line) for line in annotations.read_text(encoding="utf-8").splitlines() if line]
final=[];public=[];private=[]
for record in records:
metadata=json.dumps({key:record.get(key) for key in META_KEYS},ensure_ascii=False,separators=(",",":"));image=Path(record["image_path"]).name
for question in record["questions"]:
base={"dataset_version":record["dataset_version"],"question_id":question["question_id"],"task":TASKS[question["difficulty_level"]],"image":image,"prompt":question["question_text"]};groundtruth=str(question["ground_truth"])
public.append(base);private.append({**base,"groundtruth":groundtruth,"answer_format":question["answer_format"]});final.append({"dataset_version":record["dataset_version"],"task":base["task"],"image":image,"prompt":base["prompt"],"groundtruth":groundtruth,"metadata":metadata})
def write(name,fields,rows):
with (root/name).open("w",encoding="utf-8-sig",newline="") as handle:
writer=csv.DictWriter(handle,fieldnames=fields,lineterminator="\n");writer.writeheader();writer.writerows(rows)
write("question_set.csv",["dataset_version","question_id","task","image","prompt"],public)
write("answer_key.csv",["dataset_version","question_id","task","image","prompt","groundtruth","answer_format"],private)
write("dataset_final.csv",["dataset_version","task","image","prompt","groundtruth","metadata"],final)
with (root/"dataset_final.jsonl").open("w",encoding="utf-8",newline="\n") as handle:
for row in final:handle.write(json.dumps(row,ensure_ascii=False,separators=(",",":"))+"\n")
if len(final)!=5*len(records):raise ValueError("v7 flattened row count mismatch")
for path in (root/"question_set.csv",root/"answer_key.csv",root/"dataset_final.csv",root/"dataset_final.jsonl"):
text=path.read_text(encoding="utf-8-sig")
if "offset_intended" in text or "offset_requested_pre_containment" in text:raise ValueError(f"private offset field leaked into {path.name}")
print(f"Images: {len(records)}\nRows written: {len(final)}\nSanity check: PASS")
|