yujunjun/olmocr
1
0
Fork 0
mirror of https://github.com/allenai/olmocr.git synced 2025-10-13 09:12:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' of https://github.com/allenai/olmocr

Browse Source
This commit is contained in:
Jake Poznanski 2025-04-07 20:27:35 +00:00
commit b97e90ce3a
2 changed files with 34 additions and 65 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

53
olmocr/bench/runners/run_gemini.py
View File

@ -1,8 +1,8 @@
import base64 import base64
import os import os
from google.ai import generativelanguage as glm from google import genai
from google.api_core import client_options from google.genai import types
from olmocr.data.renderpdf import render_pdf_to_base64png from olmocr.data.renderpdf import render_pdf_to_base64png
from olmocr.prompts.anchor import get_anchor_text from olmocr.prompts.anchor import get_anchor_text
@ -30,56 +30,25 @@ def run_gemini(pdf_path: str, page_num: int = 1, model: str = "gemini-2.0-flash"
image_base64 = render_pdf_to_base64png(pdf_path, page_num=page_num, target_longest_image_dim=2048) image_base64 = render_pdf_to_base64png(pdf_path, page_num=page_num, target_longest_image_dim=2048)
anchor_text = get_anchor_text(pdf_path, page_num, pdf_engine="pdfreport") anchor_text = get_anchor_text(pdf_path, page_num, pdf_engine="pdfreport")
api_key = os.getenv("GEMINI_API_KEY") api_key = os.getenv("GEMINI_API_KEY")
client = glm.GenerativeServiceClient( client = genai.Client(api_key=api_key)
client_options=client_options.ClientOptions( image_part = types.Part(inline_data=types.Blob(mime_type="image/png", data=base64.b64decode(image_base64)))
api_key=api_key,
),
)
image_part = glm.Part(inline_data=glm.Blob(mime_type="image/png", data=base64.b64decode(image_base64))) text_part = types.Part(text=f"""{build_openai_silver_data_prompt(anchor_text)}""")
generation_config = types.GenerateContentConfig(
text_part = glm.Part(text=f"""{build_openai_silver_data_prompt(anchor_text)}""")
generation_config = glm.GenerationConfig(
temperature=temperature, temperature=temperature,
top_p=1.0, top_p=1.0,
top_k=32, top_k=32,
max_output_tokens=4096, max_output_tokens=4096,
) )
# response_schema = gemini_response_format_schema()
request = glm.GenerateContentRequest( response = client.models.generate_content(
model=f"models/{model}", model=f"models/{model}",
contents=[glm.Content(parts=[image_part, text_part])], contents=[types.Content(parts=[image_part, text_part])],
generation_config=generation_config, config=generation_config,
) )
# request = glm.GenerateContentRequest(
# model=f"models/{model}",
# contents=[glm.Content(parts=[image_part, text_part])],
# generation_config=generation_config,
# tools=[
# glm.Tool(
# function_declarations=[
# glm.FunctionDeclaration(
# name="page_response",
# parameters=response_schema
# )
# ]
# )
# ],
# tool_config=glm.ToolConfig(
# function_calling_config=glm.FunctionCallingConfig(
# mode="any",
# allowed_function_names=["page_response"]
# )
# )
# )
response = client.generate_content(request)
assert len(response.candidates) > 0, "No candidates found" assert len(response.candidates) > 0, "No candidates found"
assert ( assert response.candidates[0].finish_reason == types.FinishReason.STOP, "Finish reason was not STOP, likely a processing error or repetition failure"
response.candidates[0].finish_reason == glm.Candidate.FinishReason.STOP
), "Finish reason was not STOP, likely a processing error or repetition failure"
result = response.candidates[0].content.parts[0].text result = response.candidates[0].content.parts[0].text
return result return result

44
olmocr/bench/scripts/convert_all.sh
View File

@ -158,12 +158,12 @@ source activate olmocr
# Run olmocr benchmarks, exactly as the pipeline.py does it # Run olmocr benchmarks, exactly as the pipeline.py does it
echo "Running olmocr benchmarks..." echo "Running olmocr benchmarks..."
python -m olmocr.bench.convert olmocr_pipeline --parralel 50 python -m olmocr.bench.convert olmocr_pipeline --parallel 50 --dir olmOCR-bench/bench_data/
# Install marker-pdf and run benchmarks # Install marker-pdf and run benchmarks
echo "Installing marker-pdf and running benchmarks..." echo "Installing marker-pdf and running benchmarks..."
pip install marker-pdf==1.6.1 pip install marker-pdf==1.6.1
python -m olmocr.bench.convert marker python -m olmocr.bench.convert marker --dir olmOCR-bench/bench_data/
# Install verovio and run benchmarks # Install verovio and run benchmarks
# echo "Installing verovio and running benchmarks..." # echo "Installing verovio and running benchmarks..."
@ -172,12 +172,12 @@ python -m olmocr.bench.convert marker
# Run chatgpt benchmarks # Run chatgpt benchmarks
echo "Running chatgpt benchmarks..." echo "Running chatgpt benchmarks..."
python -m olmocr.bench.convert chatgpt python -m olmocr.bench.convert chatgpt --dir olmOCR-bench/bench_data/
#python -m olmocr.bench.convert chatgpt:name=chatgpt45:model=gpt-4.5-preview-2025-02-27 #python -m olmocr.bench.convert chatgpt:name=chatgpt45:model=gpt-4.5-preview-2025-02-27
# Run gemini benchmarks # Run gemini benchmarks
echo "Running gemini benchmarks..." echo "Running gemini benchmarks..."
python -m olmocr.bench.convert gemini:name=gemini_flash2:model=gemini-2.0-flash --parallel 4 python -m olmocr.bench.convert gemini:name=gemini_flash2:model=gemini-2.0-flash --parallel 4 --dir olmOCR-bench/bench_data/
echo "Running mistral..." echo "Running mistral..."
pip install mistralai pip install mistralai
@ -190,27 +190,27 @@ python -m olmocr.bench.convert --dir olmOCR-bench/bench_data mistral
check_port || exit 1 check_port || exit 1
# olmocr_base_temp0_1 using sglang server # olmocr_base_temp0_1 using sglang server
start_server sglang "allenai/olmOCR-7B-0225-preview" --chat-template qwen2-vl --mem-fraction-static 0.7 # start_server sglang "allenai/olmOCR-7B-0225-preview" --chat-template qwen2-vl --mem-fraction-static 0.7
python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_0:model=allenai/olmOCR-7B-0225-preview:temperature=0.0:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50 # python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_0:model=allenai/olmOCR-7B-0225-preview:temperature=0.0:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50
python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_1:model=allenai/olmOCR-7B-0225-preview:temperature=0.1:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50 # python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_1:model=allenai/olmOCR-7B-0225-preview:temperature=0.1:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50
python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_2:model=allenai/olmOCR-7B-0225-preview:temperature=0.2:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50 # python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_2:model=allenai/olmOCR-7B-0225-preview:temperature=0.2:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50
python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_3:model=allenai/olmOCR-7B-0225-preview:temperature=0.3:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50 # python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_3:model=allenai/olmOCR-7B-0225-preview:temperature=0.3:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50
python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_4:model=allenai/olmOCR-7B-0225-preview:temperature=0.4:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50 # python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_4:model=allenai/olmOCR-7B-0225-preview:temperature=0.4:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50
python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_5:model=allenai/olmOCR-7B-0225-preview:temperature=0.5:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50 # python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_5:model=allenai/olmOCR-7B-0225-preview:temperature=0.5:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50
python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_6:model=allenai/olmOCR-7B-0225-preview:temperature=0.6:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50 # python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_6:model=allenai/olmOCR-7B-0225-preview:temperature=0.6:prompt_template=fine_tune:response_template=json --repeats 1 --parallel 50
python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_7:model=allenai/olmOCR-7B-0225-preview:temperature=0.7:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50 # python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp0_7:model=allenai/olmOCR-7B-0225-preview:temperature=0.7:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50
python -m olmocr.bench.convert server:name=olmocr_base_temp0_1:model=allenai/olmOCR-7B-0225-preview:temperature=0.1:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50 # python -m olmocr.bench.convert server:name=olmocr_base_temp0_1:model=allenai/olmOCR-7B-0225-preview:temperature=0.1:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50
python -m olmocr.bench.convert server:name=olmocr_base_temp0_8:model=allenai/olmOCR-7B-0225-preview:temperature=0.8:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50 # python -m olmocr.bench.convert server:name=olmocr_base_temp0_8:model=allenai/olmOCR-7B-0225-preview:temperature=0.8:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50
stop_server # stop_server
start_server vllm "allenai/olmOCR-7B-0225-preview" # start_server vllm "allenai/olmOCR-7B-0225-preview"
python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp_vllm0_1:model=allenai/olmOCR-7B-0225-preview:temperature=0.1:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50 # python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp_vllm0_1:model=allenai/olmOCR-7B-0225-preview:temperature=0.1:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50
python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp_vllm0_7:model=allenai/olmOCR-7B-0225-preview:temperature=0.7:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50 # python -m olmocr.bench.convert --dir olmOCR-bench/bench_data server:name=olmocr_base_temp_vllm0_7:model=allenai/olmOCR-7B-0225-preview:temperature=0.7:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50
python -m olmocr.bench.convert server:name=olmocr_base_vllm_temp0_1:model=allenai/olmOCR-7B-0225-preview:temperature=0.1:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50 # python -m olmocr.bench.convert server:name=olmocr_base_vllm_temp0_1:model=allenai/olmOCR-7B-0225-preview:temperature=0.1:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50
python -m olmocr.bench.convert server:name=olmocr_base_vllm_temp0_8:model=allenai/olmOCR-7B-0225-preview:temperature=0.8:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50 # python -m olmocr.bench.convert server:name=olmocr_base_vllm_temp0_8:model=allenai/olmOCR-7B-0225-preview:temperature=0.8:prompt_template=fine_tune:response_template=json --repeats 5 --parallel 50
stop_server # stop_server
# Feel free to enable if you want # Feel free to enable if you want
# qwen2_vl_7b using sglang server # qwen2_vl_7b using sglang server