Datasetpublic

Siva2022/esg-extractor-design-and-code

ESG Metric Extractor — Design & Code Package Two files, both copy-paste ready: File Contents DESIGN.md Full design: task framing, multimodal architecture, model choices with 2026 costs, data strategy, training config (TRL-grounded), evaluation, risks, roadmap CODE.md All runnable Colab cells: Part A = v1 text-only pipeline (Qwen2.5-3B QLoRA, data prep, training, eval); Part B = v2 multimodal pipeline (Qwen3-VL-4B QLoRA on page images, teacher labeling, PDF pipeline)… See the full description on the dataset page: https://huggingface.co/datasets/Siva2022/esg-extractor-design-and-code.

sourceHugging Faceupdated 29d agoView on Hugging Face

0likes60downloads

Dataset Card

ESG Metric Extractor — Design & Code Package

Two files, both copy-paste ready:

File	Contents
DESIGN.md	Full design: task framing, multimodal architecture, model choices with 2026 costs, data strategy, training config (TRL-grounded), evaluation, risks, roadmap
CODE.md	All runnable Colab cells: Part A = v1 text-only pipeline (Qwen2.5-3B QLoRA, data prep, training, eval); Part B = v2 multimodal pipeline (Qwen3-VL-4B QLoRA on page images, teacher labeling, PDF pipeline)

Quick start (Colab)

1.Runtime → Change runtime type → L4 GPU.
2.Open CODE.md, run Part A cells in order (A1→A4 = setup + data prep; check the metric distribution output; A5→A6 = training, ~1–2 h).
3.Part B (v2) requires labeled pages of your own reports — teacher labeling (paid HF Inference credits) or a manual spreadsheet.
4.Everything pushes to your Hub namespace: Siva2022/esg-metric-extractor-qwen2.5-3b (v1) and Siva2022/esg-vlm-extractor-qwen3vl-4b (v2).

Data sources

—Beck et al. (2025) gold GHG dataset, Zenodo 14035800 (handled automatically by A3).
—Your own uploaded ESG report PDFs (real-world experiment material for Part B).

Grounding

Training configuration is taken from TRL's own verified examples (huggingface/trl @ f622980: examples/sft_qwen3_vl/sft_qwen3_vl.ipynb), not from memory.