Add app with hf inference

README.md

@@ -1,6 +1,6 @@
 ---
 title: Emission Extractor
-emoji: 🐨
+emoji: 💨
 colorFrom: gray
 colorTo: green
 sdk: gradio
@@ -8,6 +8,21 @@ sdk_version: 4.16.0
 app_file: app.py
 pinned: false
 license: apache-2.0
+preload_from_hub:
+  - mistralai/Mistral-7B-Instruct-v0.2
+  - nopperl/emissions-extraction-lora
+datasets:
+  - nopperl/sustainability-report-emissions-instruction-style
+  - nopperl/corporate-emission-reports
+tags:
+  - information-extraction
+  - retrieval
+  - climate
+  - sustainability-reports
+  - corporate-social-responsibility
+  - emissions
+  - greenhouse-gas-emissions
+  - co2-emissions
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,32 @@
+import gradio as gr
+
+from corporate_emission_reports.inference import extract_emissions
+
+
+def predict(input_method, document_file, document_url):
+    document_path = document_file if input_method == "File" else document_url
+    emissions = extract_emissions(document_path, "mistralai/Mistral-7B-Instruct-v0.2", lora="nopperl/emissions-extraction-lora", engine="hf", low_cpu_mem_usage=True)
+    return emissions.model_dump_json()
+
+with open("description.md", "r") as f:
+    description = f.read().strip()
+
+with open("article.md", "r") as f:
+    article = f.read().strip()
+
+interface = gr.Interface(
+        predict,
+        inputs=[gr.Radio(choices=["File", "URL"], value="File"), gr.File(type="filepath", file_types=[".pdf"], file_count="single", label="Report File"), gr.Textbox(label="Report URL")],
+        outputs=gr.JSON(),
+        description=description,
+        examples = [
+            ["URL", None, "https://www.bms.com/assets/bms/us/en-us/pdf/bmy-2022-esg-report.pdf"],
+            ["URL", None, "https://www.7andi.com/library/dbps_data/_template_/_res/en/sustainability/sustainabilityreport/2022/pdf/2022_all_01.pdf"],
+            ["URL", None, "https://www.infineon.com/dgdl/Sustainability_at+Infineon_2023.pdf?fileId=8ac78c8b8b657de2018c009d03120100"],
+        ],
+        article=article,
+        analytics_enabled=False,
+        cache_examples=False,
+    )
+interface.queue().launch(debug=True, share=True)
+

article.md

@@ -0,0 +1 @@
+Technical overview: The system retrieves the relevant pages of the uploaded report using simple search. These pages are input into a finetuned [Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2) language model, which outputs a JSON object containing the emission information. The system achieves an emission extraction accuracy of 62% and a source citation accuracy of 67% on the [corporate-emission-reports](https://huggingface.co/datasets/nopperl/corporate-emission-reports) dataset. Note that the model is quantized due to resource limitations.

description.md

@@ -0,0 +1,3 @@
+Upload (or link to an URL of) a sustainability report PDF and get the most recent scope 1, 2 and 3 greenhouse gas emissions reported in the document. Since this space does not use a GPU, it will be quite slow. The system was developed for English documents.
+
+Notes about the output: the emission values are in metric tons of CO2eq. The `sources` field is a list of pages containing the emission values. The page numbers are 0-based, e.g. the number 12 will correspond to page number 13 in your PDF viewer.

requirements.txt

@@ -0,0 +1 @@
+git+https://github.com/nopperl/corporate_emission_reports