wony617
Support smolagent (#5)
eaea5aa unverified
raw
history blame
8.79 kB
import os
import re
import string
import requests
from langchain.callbacks import get_openai_callback
from langchain_anthropic import ChatAnthropic
import boto3
import json
from translator.prompt_glossary import PROMPT_WITH_GLOSSARY
from translator.project_config import get_project_config
def get_content(filepath: str, project: str = "transformers") -> str:
if filepath == "":
raise ValueError("No files selected for translation.")
config = get_project_config(project)
# Extract repo path from repo_url (e.g., "huggingface/transformers")
repo_path = config.repo_url.replace("https://github.com/", "")
url = f"https://raw.githubusercontent.com/{repo_path}/main/{filepath}"
response = requests.get(url)
if response.status_code == 200:
content = response.text
return content
else:
raise ValueError("Failed to retrieve content from the URL.", url)
def preprocess_content(content: str) -> str:
# Extract text to translate from document
## ignore top license comment
to_translate = content[content.find("#") :]
## remove code blocks from text
# to_translate = re.sub(r"```.*?```", "", to_translate, flags=re.DOTALL)
## remove markdown tables from text
# to_translate = re.sub(r"^\|.*\|$\n?", "", to_translate, flags=re.MULTILINE)
## remove empty lines from text
to_translate = re.sub(r"\n\n+", "\n\n", to_translate)
return to_translate
def get_full_prompt(language: str, to_translate: str, additional_instruction: str = "") -> str:
base_prompt = string.Template(
"What do these sentences about Hugging Face Transformers "
"(a machine learning library) mean in $language? "
"Please do not translate the word after a 🤗 emoji "
"as it is a product name. Output the complete markdown file**, with prose translated and all other content intact"
"No explanations or extras—only the translated markdown. Also translate all comments within code blocks as well."
).safe_substitute(language=language)
base_prompt += "\n\n```md"
full_prompt = "\n".join([base_prompt, to_translate.strip(), "```", PROMPT_WITH_GLOSSARY])
if additional_instruction.strip():
full_prompt += f"\n\n🗒️ Additional instructions: {additional_instruction.strip()}"
return full_prompt
def split_markdown_sections(markdown: str) -> list:
# Find all titles using regular expressions
return re.split(r"^(#+\s+)(.*)$", markdown, flags=re.MULTILINE)[1:]
# format is like [level, title, content, level, title, content, ...]
def get_anchors(divided: list) -> list:
anchors = []
# from https://github.com/huggingface/doc-builder/blob/01b262bae90d66e1150cdbf58c83c02733ed4366/src/doc_builder/build_doc.py#L300-L302
for title in divided[1::3]:
anchor = re.sub(r"[^a-z0-9\s]+", "", title.lower())
anchor = re.sub(r"\s{2,}", " ", anchor.strip()).replace(" ", "-")
anchors.append(f"[[{anchor}]]")
return anchors
def make_scaffold(content: str, to_translate: str) -> string.Template:
scaffold = content
for i, text in enumerate(to_translate.split("\n\n")):
scaffold = scaffold.replace(text, f"$hf_i18n_placeholder{i}", 1)
print("inner scaffold:")
print(scaffold)
return string.Template(scaffold)
def is_in_code_block(text: str, position: int) -> bool:
"""Check if a position in text is inside a code block"""
text_before = text[:position]
code_block_starts = text_before.count("```")
return code_block_starts % 2 == 1
def fill_scaffold(content: str, to_translate: str, translated: str) -> str:
scaffold = make_scaffold(content, to_translate)
print("scaffold:")
print(scaffold.template)
# Get original text sections to maintain structure
original_sections = to_translate.split("\n\n")
# Split markdown sections to get headers and anchors
divided = split_markdown_sections(to_translate)
print("divided:")
print(divided)
anchors = get_anchors(divided)
# Split translated content by markdown sections
translated_divided = split_markdown_sections(translated)
print("translated divided:")
print(translated_divided)
# Ensure we have the same number of headers as the original
if len(translated_divided[1::3]) != len(anchors):
print(f"Warning: Header count mismatch. Original: {len(anchors)}, Translated: {len(translated_divided[1::3])}")
# Adjust anchors list to match translated headers
if len(translated_divided[1::3]) < len(anchors):
anchors = anchors[:len(translated_divided[1::3])]
else:
# Add empty anchors for extra headers
anchors.extend([""] * (len(translated_divided[1::3]) - len(anchors)))
# Add anchors to translated headers only if they're not in code blocks
for i, korean_title in enumerate(translated_divided[1::3]):
if i < len(anchors):
# Find the position of this header in the original translated text
header_pos = translated.find(korean_title.strip())
if header_pos != -1 and not is_in_code_block(translated, header_pos):
translated_divided[1 + i * 3] = f"{korean_title} {anchors[i]}"
else:
translated_divided[1 + i * 3] = korean_title
# Reconstruct translated content with proper structure
reconstructed_translated = "".join([
"".join(translated_divided[i * 3 : i * 3 + 3])
for i in range(len(translated_divided) // 3)
])
# Split by double newlines to match original structure
translated_sections = reconstructed_translated.split("\n\n")
print("scaffold template count:")
print(scaffold.template.count("$hf_i18n_placeholder"))
print("original sections length:")
print(len(original_sections))
print("translated sections length:")
print(len(translated_sections))
# Ensure section counts match
placeholder_count = scaffold.template.count("$hf_i18n_placeholder")
if len(translated_sections) < placeholder_count:
# Add empty sections if translated has fewer sections
translated_sections.extend([""] * (placeholder_count - len(translated_sections)))
elif len(translated_sections) > placeholder_count:
# Truncate if translated has more sections
translated_sections = translated_sections[:placeholder_count]
# Final check
if len(translated_sections) != placeholder_count:
return f"Error: Section count mismatch. Expected: {placeholder_count}, Got: {len(translated_sections)}"
translated_doc = scaffold.safe_substitute(
{f"hf_i18n_placeholder{i}": text for i, text in enumerate(translated_sections)}
)
return translated_doc
def llm_translate(to_translate: str) -> tuple[str, str]:
anthropic_api_key = os.environ.get("ANTHROPIC_API_KEY")
aws_bearer_token_bedrock = os.environ.get("AWS_BEARER_TOKEN_BEDROCK")
if anthropic_api_key:
# Use Anthropic API Key
model = ChatAnthropic(
model="claude-sonnet-4-20250514", max_tokens=64000, streaming=True
)
ai_message = model.invoke(to_translate)
cb = "Anthropic API Key used"
return str(cb), ai_message.content
elif aws_bearer_token_bedrock:
# Use AWS Bedrock with bearer token (assuming standard AWS credential chain is configured)
# Note: boto3 does not directly use a 'bearer_token' named environment variable for SigV4 authentication.
# It relies on AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN, or IAM roles.
# If AWS_BEARER_TOKEN_BEDROCK is meant to be one of these, it should be renamed accordingly.
# For now, we proceed assuming standard AWS credential chain is configured to pick up credentials.
client = boto3.client("bedrock-runtime", region_name="eu-north-1")
body = {
"messages": [
{"role": "user", "content": to_translate}
],
"max_tokens": 128000,
"anthropic_version": "bedrock-2023-05-31"
}
response = client.invoke_model(
modelId="arn:aws:bedrock:eu-north-1:235729104418:inference-profile/eu.anthropic.claude-3-7-sonnet-20250219-v1:0",
contentType="application/json",
accept="application/json",
body=json.dumps(body),
)
result = json.loads(response["body"].read())
cb = result["usage"]
ai_message = result["content"][0]["text"]
return str(cb), ai_message
else:
raise ValueError("No API key found for translation. Please set ANTHROPIC_API_KEY or AWS_BEARER_TOKEN_BEDROCK environment variable.")