EF Poles Classifier
Fine-tuned Qwen3-0.6B via GRPO for classifying EU development projects to Expertise France thematic poles and transversal axes.
Input: Project description + DAC codes + country
Output: List of EF poles (1–22) + transversal axes (1–6)
Reward: Composite — 0.2 × clean termination + 0.8 × semantic score, where semantic score = 0.1 (format bonus) + 0.9 × Jaccard similarity over a unified 28-class space (poles 1–22, transversal axes encoded as 23–28)
Training Data
The model was trained on a synthetic dataset (JZSG/ef_training_datasets) generated with Gemini 3.0 Flash from 1 442 original labeled examples, using a four-phase pipeline:
- Rebalancing of under-represented poles
- Variations on existing examples
- Country diversification
- Edge cases and ambiguous assignments
The synthetic dataset was split into train / test sets. Training was done on the train split; all evaluation figures below are on the held-out test split.
| Split | Examples |
|---|---|
| Train | 10,295 |
| Test | 1,144 |
Training Setup
| Parameter | Value |
|---|---|
| Base model | Qwen3-0.6B |
| Method | GRPO |
| Infrastructure | Jean Zay (IDRIS) |
| Temperature | 1.0 |
Training pipelines and evaluation scripts: Pleias/EF_training (private).
- Downloads last month
- 1