POntAvignon (Policy Ontology for Avignon festival) is small reasoning model based on Pleias-350m, specialized for the transformation of unstructured theatre program into structured semantic data.
Training
POntAvignon is based on variant of Pleias-350m mid-trained on Wikidata with generalist abilities for the transformation of unstructured text into knowledge graphs.
The model has been further trained on samples of Avignon show programs with a reinforcement learning method (GRPO). We initially focused on constraining the model to only generate items about shows and use a limited set of properties. A judge model (Gemma 12B) contributed to enhance the overall accuracy of the results by assessing the preliminary drafts at each step.
Use
PontAvignon is a specialized reasoning model using a special tokens to encode inputs, structured data outputs and the intermediary reasoning traces.
Any excerpts of a show program has to be submitted using the tags '<|text_startl>' and '<|text_end|>'.
After '<|text_end|>' will model will generate an answer in three step:
- Initial fuzzy "thinking" (between <|thinking_start|> and <|thinking_end|>):
Some warnings apply:
- Given the intensive reinforcement learning specialization, the model will mostly work on theater show in French, preferably one structured similarly to original sources from the festival d'Avignon. More diverse sources would be needed to make the model more source agnostic.
- To improve results accuracy, we trained the model on filtered versions of the original sources with all the information about a unique show. Multi-show submissions will either result on the model focusing on only one show or, even more problematic, mixing information.
- Reasoning traces contribute to increase the accuracy of the model and provide some level of explainability. They can still be counterintuitive and potentially contain wrong assumptions that will not necessarily be kept in the final output.
