DReggio/mrbert-es-esg

MrBERT-es ESG News Classifier (Spanish)

Model summary

Fine-tuned ESG news classifier for Spanish equity market headlines. Based on BSC-LT/MrBERT-es (ModernBERT, 150M parameters, bilingual ES/EN). Classifies Spanish financial news headlines into ESG pillars (Environmental, Social, Governance) and sentiment (Positive / Negative / Neutral / NA). Training regime: human-gold annotations (1,688 events) augmented with LLaMA-3.1-8B SFT silver labels (62,800 events).

Repository layout

This repository ships four sibling sub-models — one per task head — because the fine-tuned architecture is a shared BSC-LT/MrBERT-es encoder with separate classification heads (three binary pillar heads plus one 4-class sentiment head), not a single multi-class classifier.

mrbert-es-esg/
├── esg_E/         binary head: Environmental pillar (0 = not-E, 1 = E)
├── esg_S/         binary head: Social pillar       (0 = not-S, 1 = S)
├── esg_G/         binary head: Governance pillar   (0 = not-G, 1 = G)
└── sentiment/     4-class head: {Pos, Neg, Neu, NA}

Each sub-folder contains:

• model.safetensors — TAPT-adapted ModernBERT-es encoder (~599 MB)
• separate_head.pt — classification head weights for this task
• config.json — ModernBertModel encoder config
• separate_classifier_config.json — head metadata (pillar, hidden_size, num_sentiment_classes)
• tokenizer.json, tokenizer_config.json — fast tokenizer (ModernBERT uses a single tokenizer.json; no vocab.txt / special_tokens_map.json)

ESG pillar predictions are independent binary classifications — the same headline can be flagged on any combination of E, S, G, or none (multi-label by design, per the project codebook).

Intended use

• ESG signal extraction from Spanish business press
• Event study research on ESG news and equity market response
• Spanish financial NLP benchmarking

Out of scope: high-stakes automated decisions without human review; languages other than Spanish.

How to use

The classification head is a custom nn.Module stored in separate_head.pt, separate from the encoder weights — so the standard AutoModelForSequenceClassification / pipeline("text-classification", ...) path does not work out of the box. To run inference, load the encoder with AutoModel and apply the matching separate_head.pt for each task.

Minimum loading sketch (encoder only — head loading uses the project's SeparateClassifier class):

import torch
from transformers import AutoTokenizer, AutoModel

head = "esg_E"  # or "esg_S", "esg_G", "sentiment"
tok = AutoTokenizer.from_pretrained(f"DReggio/mrbert-es-esg/{head}")
enc = AutoModel.from_pretrained(f"DReggio/mrbert-es-esg/{head}")
state = torch.load(f"DReggio/mrbert-es-esg/{head}/separate_head.pt")
# instantiate SeparateClassifier(...) from project repo, load state, forward.

Training data

Gold set: 1,688 human-annotated Spanish ESG events (E / S / G binary + 4-class sentiment). Silver augmentation (R3): 62,800 events labelled by a LLaMA-3.1-8B SFT annotator (variant4 prompt, masked-loss QLoRA, mean κ = 0.761 vs human gold). R3 was selected over R2 (Qwen3-8B SFT silver) because LLaMA's higher Governance base-rate yields richer positive-class signal for the G pillar; the R3 fine-tune lifts macro-F1 by 0.7–3.1 pp and MCC_G by 2.2–6.6 pp over the gold-only R1b baseline.

Evaluation (gold_test, n = 267)

Metric	Value
Macro F1	0.8460
F1_E	0.8866
F1_S	0.8485
F1_G	0.8029
MCC_G	0.5953
κ_sentiment	0.7006

Training procedure

• Base model: BSC-LT/MrBERT-es
• Architecture: separate binary classification head per pillar (E / S / G)
  • 4-class sentiment head, over a shared TAPT-adapted ModernBERT-es encoder
• Training regimes: R1b (gold only) → R2 (gold + Qwen3 silver) → R3 (gold + LLaMA silver, this checkpoint)
• Framework: HuggingFace Transformers + PyTorch
• Hardware: Google Colab Pro A100 40GB

Limitations

• Governance (G) pillar F1 0.80, MCC_G 0.60 — structurally ambiguous category; use with caution for G-specific downstream tasks.
• Silver-label bias: R3 inherits the labelling distribution of the LLaMA-3.1-8B SFT annotator, which has a higher G base-rate (47.5 %) than the Qwen3 silver corpus (34.0 %); applications sensitive to G prevalence should consider the gold-only R1b checkpoint or the Qwen3-silver R2 checkpoint as sensitivity comparisons.
• Training data: Spanish peninsular financial press (2014–2024); Latin American Spanish and social media not covered.

License

Apache 2.0 — derived from BSC-LT/MrBERT-es (© Barcelona Supercomputing Center, Apache 2.0).

Citation

@mastersthesis{reggio2026esg,
  author = {Damien Reggio},
  title  = {ESG News Classification and Market Response
            in Spanish Equity Markets},
  school = {FernUni Switzerland},
  year   = {2026}