done

Author: Hangzhi

dspy/predict/chain_of_thought_with_hint.py

@@ -0,0 +1,28 @@
+import dspy
+
+from .predict import Module
+
+
+class ChainOfThoughtWithHint(Module):
+    def __init__(self, signature, rationale_field_type=None, **config):
+        self.signature = dspy.ensure_signature(signature)
+        self.module = dspy.ChainOfThought(signature, rationale_field_type=rationale_field_type, **config)
+
+    def forward(self, **kwargs):
+        if kwargs.get("hint"):
+            hint = f"\n\t\t(secret hint: {kwargs.pop('hint')})"
+            original_kwargs = kwargs.copy()
+
+            # Convert the first field's value to string and append the hint
+            last_key = list(self.signature.input_fields.keys())[-1]
+            kwargs[last_key] = str(kwargs[last_key]) + hint
+
+            # Run CoT then update the trace with original kwargs, i.e. without the hint.
+            with dspy.context(trace=[]):
+                pred = self.module(**kwargs)
+                this_trace = dspy.settings.trace[-1]
+
+            dspy.settings.trace.append((this_trace[0], original_kwargs, this_trace[2]))
+            return pred
+
+        return self.module(**kwargs)

dspy/teleprompt/bootstrap_finetune.py

@@ -81,17 +81,18 @@ def compile(
         key_to_data = {}
         for pred_ind, pred in enumerate(student.predictors()):
             data_pred_ind = None if self.multitask else pred_ind
-            training_key = (pred.lm, data_pred_ind)
+            lm = pred.lm or settings.lm
+            training_key = (lm, data_pred_ind)
             if training_key not in key_to_data:
                 train_data, data_format = self._prepare_finetune_data(
-                    trace_data=trace_data, lm=pred.lm, pred_ind=data_pred_ind
+                    trace_data=trace_data, lm=lm, pred_ind=data_pred_ind
                 )
-                logger.info(f"Using {len(train_data)} data points for fine-tuning the model: {pred.lm.model}")
+                logger.info(f"Using {len(train_data)} data points for fine-tuning the model: {lm.model}")
                 finetune_kwargs = {
-                    "lm": pred.lm,
+                    "lm": lm,
                     "train_data": train_data,
                     "train_data_format": data_format,
-                    "train_kwargs": self.train_kwargs[pred.lm],
+                    "train_kwargs": self.train_kwargs[lm],
                 }
                 key_to_data[training_key] = finetune_kwargs