---
license: mit
language:
  - en
pipeline_tag: text-classification
tags:
  - nouns-dao
  - governance
  - ethereum
  - dao
  - proposals
widget:
  - text: "🎨 Noundry: Add Messi Accessory — adds a new CC0 accessory trait to the Nouns art repository"
  - text: "Proposal to stake 5000 ETH in Lido to earn yield on the treasury"
  - text: "Appoint Gramajo as DUNA Compliance Administrator for the DAO reserve"
---

# Nouns DAO Proposal Predictor

A DistilBERT model fine-tuned on [Nouns DAO Proposals](https://huggingface.co/datasets/gramajo/nouns-proposals) to predict whether a governance proposal will pass or fail based on its title and description.

## Model Description

- **Model:** `distilbert-base-uncased` (66M parameters)
- **Task:** Binary text classification (Pass / Fail)
- **Training data:** 785 Nouns DAO proposals (chronological train split)
- **Evaluation data:** 197 proposals (most recent 20%, chronological test split)
- **Training hardware:** Google Colab T4 GPU (~5 minutes)

## Performance

Evaluated on the chronological test set (proposals 786–982, where the DAO was significantly more conservative):

| Metric | Score |
|--------|-------|
| Accuracy | ~58% |
| F1 (binary) | ~0.42 |
| Precision | ~0.35 |
| Recall | ~0.54 |

**⚠️ Important context:** The test set has a 28% pass rate vs 56% in the training set due to the Nouns DAO becoming more conservative over time. This chronological split is intentional — it simulates real-world prediction of *future* proposals from historical data. A random split would show inflated ~65% F1 but would leak temporal information.

## Intended Use

- Educational tool for understanding DAO governance patterns
- Baseline model for proposal outcome prediction research
- **Not intended for:** automated voting, financial decisions, or production governance without human oversight

## Limitations

- Trained on only 785 examples — a very small dataset by ML standards
- The Nouns DAO's voting patterns, participants, and norms have shifted significantly over 5 years
- Proposal descriptions vary wildly in length and format (some have images, markdown, etc.)
- Categories (Art, Governance, Funding, Events, Infrastructure) are LLM-assigned, not human-verified
- Does not use proposal category as a feature (text-only v1)

## Training Details

- **Epochs:** 4 (with early stopping)
- **Batch size:** 16
- **Learning rate:** 2e-5
- **Max sequence length:** 512 tokens
- **Class weights:** Computed from training set imbalance (56%/44% pass/fail)
- **Loss:** Weighted cross-entropy

## Dataset

Trained on [gramajo/nouns-proposals](https://huggingface.co/datasets/gramajo/nouns-proposals) — all 983 Nouns DAO proposals from August 2021 to July 2026, with pass/fail labels and LLM-assigned categories.

## Usage

```python
from transformers import pipeline

classifier = pipeline("text-classification", model="gramajo/nouns-proposal-predictor")

result = classifier("Your proposal title and description here")
# [{'label': 'LABEL_1', 'score': 0.73}]  — LABEL_1 = Pass, LABEL_0 = Fail
```

## Citation

```bibtex
@misc{nouns-proposal-predictor-2026,
  author = {Juan Gramajo},
  title = {Nouns DAO Proposal Predictor},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\\url{https://huggingface.co/gramajo/nouns-proposal-predictor}}
}
```

## License

MIT. Built on Nouns DAO data which is CC0 (public domain).