tinker_cookbook.eval.TrainingClientEvaluator
class tinker_cookbook.eval.TrainingClientEvaluator()
An evaluator that uses a TrainingClient to compute metrics (e.g., loss).
class NLLEval(TrainingClientEvaluator):
async def __call__(self, training_client):
loss = await compute_nll(training_client, self.eval_data)
return {"eval/nll": loss}
__call__(training_client)
Run evaluation and return a dict of metric names to values.
Parameters:
- training_client (tinker.TrainingClient) – The Tinker training client to evaluate with.
Returns: A dict mapping metric names to float values.