Skip to content

FewShotOPRO

FewShotOPRO

Bases: Optimizer

Sample randomly among the best examples to populate the LM's prompt to make it learn using Few Shot Learning while generating instructions with OPRO.

Example:

import synalinks
import asyncio

async def main():
    # ... your program definition

    program.compile(
        reward=synalinks.rewards.ExactMatch(),
        optimizer=synalinks.optimizers.FewShotOPRO(
            language_model=language_model,
            k=3, # The number of examples to provide to the prompt
            k_best=10, # The number of best examples to select from
        ),
    )

    history = await program.fit(...)
References

Parameters:

Name Type Description Default
language_model LanguageModel

The language model to use.

 None
k int

The number of examples to select (default 3) among the best predictions.

 3
k_best int

The max number of best predictions/instructions to select from (default 10).

 10
program Program

The program to use. Optional. If None create one at start.

 None
name str

The name of the optimizer.

 None
description str

The description of the optimizer.

 None
Source code in synalinks/src/optimizers/few_shot_opro.py 
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
@synalinks_export("synalinks.optimizers.FewShotOPRO")
class FewShotOPRO(Optimizer):
    """Sample randomly among the best examples to populate the LM's prompt to make it
        learn using Few Shot Learning while generating instructions with OPRO.

    Example:

    ```python
    import synalinks
    import asyncio

    async def main():
        # ... your program definition

        program.compile(
            reward=synalinks.rewards.ExactMatch(),
            optimizer=synalinks.optimizers.FewShotOPRO(
                language_model=language_model,
                k=3, # The number of examples to provide to the prompt
                k_best=10, # The number of best examples to select from
            ),
        )

        history = await program.fit(...)
    ```

    References:
        - [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165)
        - [Large Language Models as Optimizers](https://arxiv.org/abs/2309.03409)

    Args:
        language_model (LanguageModel): The language model to use.
        k (int): The number of examples to select (default 3) among the best predictions.
        k_best (int): The max number of best predictions/instructions to select from
            (default 10).
        program (Program): The program to use. Optional. If None create one at start.
        name (str): The name of the optimizer.
        description (str): The description of the optimizer.
    """

    def __init__(
        self,
        language_model=None,
        k=3,
        k_best=10,
        program=None,
        name=None,
        description=None,
    ):
        super().__init__(
            name=name,
            description=description,
            data_model=FewShotOPROOptimizedVariables,
        )
        self.language_model = language_model
        self.k = k
        self.k_best = k_best
        self.program = program

    async def build(self, variables):
        if not self.program:
            opro_inputs = Input(data_model=OPROInputs)
            opro_outputs = await Generator(
                language_model=self.language_model,
                data_model=Instructions,
                instructions=[
                    "Your task is to generate instructions that maximize rewards.",
                    "Below are some previous instructions with their reward.",
                    "Generate instructions that is different from all the instructions.",
                    (
                        "The instructions should be concise, effective and generally"
                        " applicable to all predictions below."
                    ),
                ],
            )(opro_inputs)

            self.program = Program(
                inputs=opro_inputs,
                outputs=opro_outputs,
                name="opro",
                description="OPRO Program",
            )
        self.built = True

    async def optimize(self, trainable_variable, reward=None):
        """Perform a backprop/optimization on a single variable."""
        # Reward backpropagation
        predictions = trainable_variable.get("predictions")
        backpropagated_predictions = []
        backprop_pred_nb = 0
        for p in predictions:
            if p["reward"] is None:
                p["reward"] = reward
                backprop_pred_nb += 1
            backpropagated_predictions.append(p)
        if backprop_pred_nb > 0:
            trainable_variable.update({"predictions": backpropagated_predictions})
            # Get the k best predictions (sorted by reward)
            sorted_predictions = sorted(
                backpropagated_predictions,
                key=lambda x: x["reward"] if x["reward"] is not None else float("-inf"),
                reverse=True,
            )
            top_k_predictions = sorted_predictions[: self.k_best]
            if len(top_k_predictions) > self.k:
                selected_predictions = random.sample(top_k_predictions, self.k)
            else:
                selected_predictions = top_k_predictions
            # Get the k best instructions candidates (sorted by reward)
            sorted_instructions_candidates = sorted(
                trainable_variable.get("instructions_candidates"),
                key=lambda x: x["reward"] if x["reward"] is not None else float("-inf"),
                reverse=True,
            )
            top_k_instructions_candidates = sorted_instructions_candidates[: self.k_best]
            # Prepare inputs for OPRO
            inputs = OPROInputs(
                predictions=top_k_predictions,
                instructions_candidates=top_k_instructions_candidates,
            )
            new_instructions = await self.program(inputs)
            trainable_variable.update(
                {
                    "instructions": new_instructions.get_json(),
                    "examples": selected_predictions,
                }
            )

    async def finalize(self, trainable_variable):
        """Finalize the optimization of a single variable (cleanup/scaling etc.)."""
        trainable_variable.update(
            {
                "predictions": [],
                "instructions_candidates": []
            }
        )

    def get_config(self):
        config = {
            "k": self.k,
            "k_best": self.k_best,
            "name": self.name,
            "description": self.description,
        }
        language_model_config = {
            "language_model": serialization_lib.serialize_synalinks_object(
                self.language_model,
            )
        }
        program_config = {
            "program": serialization_lib.serialize_synalinks_object(
                self.program,
            )
        }
        return {**config, **language_model_config, **program_config}

    @classmethod
    def from_config(cls, config):
        language_model = serialization_lib.deserialize_synalinks_object(
            config.pop("language_model"),
        )
        program = serialization_lib.deserialize_synalinks_object(
            config.pop("program"),
        )
        return cls(language_model=language_model, program=program, **config)

finalize(trainable_variable) async

Finalize the optimization of a single variable (cleanup/scaling etc.).

Source code in synalinks/src/optimizers/few_shot_opro.py 
153
154
155
156
157
158
159
160
async def finalize(self, trainable_variable):
    """Finalize the optimization of a single variable (cleanup/scaling etc.)."""
    trainable_variable.update(
        {
            "predictions": [],
            "instructions_candidates": []
        }
    )

optimize(trainable_variable, reward=None) async

Perform a backprop/optimization on a single variable.

Source code in synalinks/src/optimizers/few_shot_opro.py 
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
async def optimize(self, trainable_variable, reward=None):
    """Perform a backprop/optimization on a single variable."""
    # Reward backpropagation
    predictions = trainable_variable.get("predictions")
    backpropagated_predictions = []
    backprop_pred_nb = 0
    for p in predictions:
        if p["reward"] is None:
            p["reward"] = reward
            backprop_pred_nb += 1
        backpropagated_predictions.append(p)
    if backprop_pred_nb > 0:
        trainable_variable.update({"predictions": backpropagated_predictions})
        # Get the k best predictions (sorted by reward)
        sorted_predictions = sorted(
            backpropagated_predictions,
            key=lambda x: x["reward"] if x["reward"] is not None else float("-inf"),
            reverse=True,
        )
        top_k_predictions = sorted_predictions[: self.k_best]
        if len(top_k_predictions) > self.k:
            selected_predictions = random.sample(top_k_predictions, self.k)
        else:
            selected_predictions = top_k_predictions
        # Get the k best instructions candidates (sorted by reward)
        sorted_instructions_candidates = sorted(
            trainable_variable.get("instructions_candidates"),
            key=lambda x: x["reward"] if x["reward"] is not None else float("-inf"),
            reverse=True,
        )
        top_k_instructions_candidates = sorted_instructions_candidates[: self.k_best]
        # Prepare inputs for OPRO
        inputs = OPROInputs(
            predictions=top_k_predictions,
            instructions_candidates=top_k_instructions_candidates,
        )
        new_instructions = await self.program(inputs)
        trainable_variable.update(
            {
                "instructions": new_instructions.get_json(),
                "examples": selected_predictions,
            }
        )