sebastien/Copies
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Miscs

Browse Source
This commit is contained in:
sebastien
2026-09-08 16:41:04 +02:00
parent db4ed2ef31
commit bf05272797
14 changed files with 725 additions and 75 deletions
Download Patch File Download Diff File Expand all files Collapse all files
copienator/commands/gemini_for_labels.py
+125 -7
View File
@@ -76,6 +76,8 @@ be missing.
##wrong_labels##
##wrong_label_text_context##
Here's a list of the names of the students, pick the one that matches
the best or `\"Unknown\"` if you cannot read the name
@@ -133,6 +135,8 @@ be missing.
##wrong_labels##
##wrong_label_text_context##
Since this copy isn't the first part of a sequence, simply set the
name to `\"Continued\"`."""
@@ -147,7 +151,66 @@ class AnnotationData(BaseModel):
)
def generate_request(file, labels, names, context_labels, wrong_labels):
TEXT_CONTEXT_MAX_CHARS = 4000
def _label_filename(path: Path) -> str:
return path.stem if path.suffix.casefold() in {".tex", ".txt"} else path.name
def _common_prefix_length(left: str, right: str) -> int:
left_folded = left.casefold()
right_folded = right.casefold()
limit = min(len(left_folded), len(right_folded))
for index in range(limit):
if left_folded[index] != right_folded[index]:
return index
return limit
def closest_text_context(
workspace: EvaluationWorkspace, wrong_labels: list[str]
) -> tuple[Path | None, str]:
"""Return a bounded excerpt from the Text file closest to an invalid label."""
text_dir = workspace.root / "Text"
if not wrong_labels or not text_dir.is_dir():
return None, ""
ranked: list[tuple[int, str, Path]] = []
for path in text_dir.iterdir():
if not path.is_file() or path.suffix.casefold() == ".pdf":
continue
filename = _label_filename(path)
prefix_length = max(
_common_prefix_length(filename, wrong_label)
for wrong_label in wrong_labels
)
if prefix_length:
ranked.append((prefix_length, filename.casefold(), path))
for _prefix_length, _filename, path in sorted(
ranked, key=lambda item: (-item[0], item[1])
):
try:
content = path.read_text(encoding="utf-8")
except (OSError, UnicodeError):
continue
if len(content) > TEXT_CONTEXT_MAX_CHARS:
content = content[:TEXT_CONTEXT_MAX_CHARS] + "\n[excerpt truncated]"
return path, content
return None, ""
def generate_request(
file,
labels,
names,
context_labels,
wrong_labels,
wrong_label_text_context="",
wrong_label_text_file: Path | None = None,
seed: int = 0,
):
"""Generates request for Gemini with context."""
image_path = Path(file)
@@ -162,9 +225,36 @@ def generate_request(file, labels, names, context_labels, wrong_labels):
text = my_prompt2.replace("##labels##", labels)\
.replace("##prev_context##", context_str)
if wrong_labels:
text= text.replace("##wrong_labels##\n\n", f"On a previous request, you answered with the following wrong labels : {wrong_labels}. These are wrong, since they do not exactly match any of the labels in the previous list.")
formatted_wrong_labels = "\n".join(f'- "{label}"' for label in wrong_labels)
text = text.replace(
"##wrong_labels##",
"On the previous request for this image, you answered with these "
"invalid labels:\n"
f"{formatted_wrong_labels}\n"
"They are wrong because they do not exactly match any label in the "
"valid list above.\n\n"
"CRITICAL RETRY CONSTRAINT: NEVER return any of the invalid labels "
"listed above again. Your answer must use only exact labels copied "
"verbatim from the valid list. If the handwriting resembles an "
"invalid label, choose the closest exact valid label instead.",
)
else:
text = text.replace("##wrong_labels##\n\n", "")
text = text.replace("##wrong_labels##", "")
if wrong_label_text_context and wrong_label_text_file:
text = text.replace(
"##wrong_label_text_context##",
"Here is an excerpt from the exam text file whose name has the "
"longest prefix in common with the invalid label(s), "
f"`{wrong_label_text_file.name}`:\n\n"
"<exam_text_excerpt>\n"
f"{wrong_label_text_context}\n"
"</exam_text_excerpt>\n\n"
"Use this excerpt as extra context for identifying the handwritten "
"label, but return only an exact label from the valid list above.",
)
else:
text = text.replace("##wrong_label_text_context##", "")
contents = [
@@ -181,9 +271,10 @@ def generate_request(file, labels, names, context_labels, wrong_labels):
]
generate_content_config = types.GenerateContentConfig(
automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
temperature=1.0,
top_p=0.95,
seed=0,
seed=seed,
max_output_tokens=65535,
response_mime_type= "application/json",
response_json_schema= AnnotationData.model_json_schema(),
@@ -293,17 +384,30 @@ def process_copy_group(
f"{len(accumulated_labels)} accumulated labels..."
)
attempt = 0
label_retry_count = 0
wrong_labels: list[str] = []
while True:
if attempt > 0:
sleep(10 * attempt)
try:
text_context_file, text_context = closest_text_context(
workspace, wrong_labels
)
if text_context_file:
print(
f"[{group_key}] Retry context for {image_file.name}: "
f"{text_context_file.relative_to(workspace.root)}"
)
request_seed = max(0, label_retry_count - 1)
contents, request_config = generate_request(
image_file,
labels_text,
names_text,
accumulated_labels,
wrong_labels,
text_context,
text_context_file,
seed=request_seed,
)
response = client.models.generate_content(
model=MODEL_ID,
@@ -321,9 +425,23 @@ def process_copy_group(
f"Error: {image_file.name} contained unknown labels: "
f"{unknown}"
)
wrong_labels.extend(unknown)
attempt += 1
continue
unique_unknown = list(dict.fromkeys(unknown))
if (
label_retry_count >= 2
and set(unique_unknown) == set(wrong_labels)
):
for item in annotation.list:
if item.label in unique_unknown:
item.label = f"??{item.label}"
print(
f"Warning: {image_file.name} repeated the same unknown "
"label(s) on the third try; keeping them with a ?? prefix."
)
else:
wrong_labels = unique_unknown
label_retry_count += 1
attempt += 1
continue
if annotation.name not in valid_names:
print(
f"Error: {image_file.name} returned unknown name: "