Classification : Implement Batch API for model evaluation

Current Behavior
For evaluation of fine-tuned models, we currently use the Responses API synchronously to generate predictions. These predictions are then compared against the true labels of the corresponding questions. However, this synchronous setup is difficult to manage and would require building a complex system to make the whole process asynchronous and more reliable.

Proposed Improvement
Instead of relying on the synchronous approach, we can leverage the OpenAI Batch API. By sending all evaluation queries together as a batch, we can retrieve the predictions from the batch output file and then compare them against the true labels. This would simplify the workflow and make the evaluation process more scalable and efficient.

document to refer : https://docs.google.com/document/d/1y2WLRUm2dZaQMetvb31XpDH5HqteRe4UNpziG8pq608/edit?tab=t.lbfxo8c09bqd#bookmark=id.1ib6z5f31fld

Examinez le flux d’évaluation synchrone actuel de la Responses API dans le dépôt, puis consultez le document de conception lié pour le workflow par lots proposé. La tâche est terminée lorsque les requêtes d’évaluation sont soumises via la Batch API, que les prédictions sont lues depuis le fichier de sortie du batch et que ces prédictions sont comparées aux vraies étiquettes correspondantes.