Skip to content

Data Mixture optimization

bolt.DMCurriculum

Data mixture selection with curriculum.

First set of data mixture is used to train from 0 to 5M tokens. Second set of data mixture is used to train from 5M to 10M tokens.

3 types of datasets are used
  • instruction following (IF): allenai/tulu-3-sft-personas-instruction-following
  • math: allenai/tulu-3-sft-personas-math
  • code: allenai/tulu-3-sft-personas-code
6 parameters
  1. IF proportion 1 (float, [0, 1])
  2. Math proportion 1 (float, [0, 1])
  3. Code proportion 1 (float, [0, 1])
  4. IF proportion 2 (float, [0, 1])
  5. Math proportion 2 (float, [0, 1])
  6. Code proportion 2 (float, [0, 1])

Parameters 1,2,3 must sum to 1. So must parameters 4,5,6. (Simplex)

Single objective that averages
  • evaluation results on IFEval (strict)
  • evaluation results on MATH-500 (minerva format)
  • evaluation results on MBPP plus

Example usage: 

import torch
from bolt import DMCurriculum

prob = DMCurriculum(noise_std=0.001)

X = torch.Tensor([[0.15, 0.2, 0.65, 0.6, 0.2, 0.2]])
y = prob(X)

Source code in bolt/problems/data_mixture.py 
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 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
class DMCurriculum(LLMTestProblem):
    r"""Data mixture selection with curriculum.

    First set of data mixture is used to train from 0 to 5M tokens.
    Second set of data mixture is used to train from 5M to 10M tokens.

    3 types of datasets are used:
        - instruction following (IF): allenai/tulu-3-sft-personas-instruction-following
        - math: allenai/tulu-3-sft-personas-math
        - code: allenai/tulu-3-sft-personas-code

    6 parameters:
        1. IF proportion 1 (float, [0, 1])
        2. Math proportion 1 (float, [0, 1])
        3. Code proportion 1 (float, [0, 1])
        4. IF proportion 2 (float, [0, 1])
        5. Math proportion 2 (float, [0, 1])
        6. Code proportion 2 (float, [0, 1])

    Parameters 1,2,3 must sum to 1. So must parameters 4,5,6. (Simplex)

    Single objective that averages:
        - evaluation results on IFEval (strict)
        - evaluation results on MATH-500 (minerva format)
        - evaluation results on MBPP plus

    Example usage:
    ```python
    import torch
    from bolt import DMCurriculum

    prob = DMCurriculum(noise_std=0.001)

    X = torch.Tensor([[0.15, 0.2, 0.65, 0.6, 0.2, 0.2]])
    y = prob(X)
    ```

    """

    name = "dm_curriculum"
    hf_repo = "chewwt/dm_qwen4b_emulator"

    dim = 6
    _bounds = [
        (0.0, 1.0),  # if_prop1
        (0.0, 1.0),  # math_prop1
        (0.0, 1.0),  # code_prop1
        (0.0, 1.0),  # if_prop2
        (0.0, 1.0),  # math_prop2
        (0.0, 1.0),  # code_prop2
    ]

    _check_grad_at_opt: bool = True
    continuous_inds = [0, 1, 2, 3, 4, 5]

    _optimal_value = 0.61116  # empirically found
    _optimizers = [(0.5663, 0.4337, 0.0000, 0.4139, 0.4901, 0.0960)]

    def __init__(
        self,
        noise_std: None | float | list[float] = None,
        negate: bool = False,
        dtype: torch.dtype = torch.double,
    ) -> None:
        r"""Data mixture curriculum optimization for Qwen3-4B-Base

        Args:
            noise_std: Standard deviation of the observation noise. If a list is
                provided, specifies separate noise standard deviations for each
                objective in a multiobjective problem.
            negate: If True, negate the function.
            dtype: The dtype that is used for the bounds of the function.
        """

        super().__init__(
            noise_std=noise_std,
            negate=negate,
            dtype=dtype,
        )

        self.model_path, self.model_config, self.y_mean, self.y_std = (
            pull_info_from_hf_hub(self.hf_repo)
        )

        self.obj_func = MLPFunction(
            self.model_config["input_dim"],
            self.model_path,
            hidden_dim=self.model_config["hidden_dim"],
            output_dim=self.model_config["output_dim"],
        )

    def _evaluate_true(self, X: torch.Tensor) -> torch.Tensor:
        r"""Evaluate the objective using the pretrained emulator.

        Args:
            X (torch.Tensor): Input tensor of shape `(N, dim)`.

        Returns:
            torch.Tensor: Objective tensor of shape `(N, 1)`.
        """

        validate_simplex_product(X=X, eps=1e-5)

        y_mo = self.obj_func.evaluate_true(X)
        y_mo_st = unstandardize_y(y_mo, self.y_mean, self.y_std)
        return y_mo_st.mean(axis=1)[..., None]

__init__(noise_std=None, negate=False, dtype=torch.double)

Data mixture curriculum optimization for Qwen3-4B-Base

Parameters:

Name Type Description Default
noise_std None | float | list[float]

Standard deviation of the observation noise. If a list is provided, specifies separate noise standard deviations for each objective in a multiobjective problem.

 None
negate bool

If True, negate the function.

 False
dtype dtype

The dtype that is used for the bounds of the function.

 double
Source code in bolt/problems/data_mixture.py 
 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
def __init__(
    self,
    noise_std: None | float | list[float] = None,
    negate: bool = False,
    dtype: torch.dtype = torch.double,
) -> None:
    r"""Data mixture curriculum optimization for Qwen3-4B-Base

    Args:
        noise_std: Standard deviation of the observation noise. If a list is
            provided, specifies separate noise standard deviations for each
            objective in a multiobjective problem.
        negate: If True, negate the function.
        dtype: The dtype that is used for the bounds of the function.
    """

    super().__init__(
        noise_std=noise_std,
        negate=negate,
        dtype=dtype,
    )

    self.model_path, self.model_config, self.y_mean, self.y_std = (
        pull_info_from_hf_hub(self.hf_repo)
    )

    self.obj_func = MLPFunction(
        self.model_config["input_dim"],
        self.model_path,
        hidden_dim=self.model_config["hidden_dim"],
        output_dim=self.model_config["output_dim"],
    )

bolt.DMCurriculumMO

Data mixture selection with curriculum and multi-objective.

First set of data mixture is used to train from 0 to 5M tokens. Second set of data mixture is used to train from 5M to 10M tokens.

3 types of datasets are used
  • instruction following (IF): allenai/tulu-3-sft-personas-instruction-following
  • math: allenai/tulu-3-sft-personas-math
  • code: allenai/tulu-3-sft-personas-code
6 parameters
  1. IF proportion 1 (float, [0, 1])
  2. Math proportion 1 (float, [0, 1])
  3. Code proportion 1 (float, [0, 1])
  4. IF proportion 2 (float, [0, 1])
  5. Math proportion 2 (float, [0, 1])
  6. Code proportion 2 (float, [0, 1])

Parameters 1,2,3 must sum to 1. So must parameters 4,5,6. (Simplex)

Multiobjective with 3 outputs
  1. evaluation results on IFEval (strict)
  2. evaluation results on MATH-500 (minerva format)
  3. evaluation results on MBPP plus

Example usage: 

import torch
from bolt import DMCurriculumMO

prob = DMCurriculumMO(noise_std=0.001)

X = torch.Tensor([[0.15, 0.2, 0.65, 0.6, 0.2, 0.2]])
y = prob(X)

Source code in bolt/problems/data_mixture.py 
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
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
class DMCurriculumMO(MultiObjectiveTestProblem, LLMTestProblem):
    r"""Data mixture selection with curriculum and multi-objective.

    First set of data mixture is used to train from 0 to 5M tokens.
    Second set of data mixture is used to train from 5M to 10M tokens.

    3 types of datasets are used:
        - instruction following (IF): allenai/tulu-3-sft-personas-instruction-following
        - math: allenai/tulu-3-sft-personas-math
        - code: allenai/tulu-3-sft-personas-code

    6 parameters:
        1. IF proportion 1 (float, [0, 1])
        2. Math proportion 1 (float, [0, 1])
        3. Code proportion 1 (float, [0, 1])
        4. IF proportion 2 (float, [0, 1])
        5. Math proportion 2 (float, [0, 1])
        6. Code proportion 2 (float, [0, 1])

    Parameters 1,2,3 must sum to 1. So must parameters 4,5,6. (Simplex)

    Multiobjective with 3 outputs:
        1. evaluation results on IFEval (strict)
        2. evaluation results on MATH-500 (minerva format)
        3. evaluation results on MBPP plus

    Example usage:
    ```python
    import torch
    from bolt import DMCurriculumMO

    prob = DMCurriculumMO(noise_std=0.001)

    X = torch.Tensor([[0.15, 0.2, 0.65, 0.6, 0.2, 0.2]])
    y = prob(X)
    ```

    """

    name = "dm_curriculum_mo"
    hf_repo = "chewwt/dm_qwen4b_emulator"

    dim = 6
    _bounds = [
        (0.0, 1.0),  # if_prop1
        (0.0, 1.0),  # math_prop1
        (0.0, 1.0),  # code_prop1
        (0.0, 1.0),  # if_prop2
        (0.0, 1.0),  # math_prop2
        (0.0, 1.0),  # code_prop2
    ]

    _check_grad_at_opt: bool = True
    continuous_inds = [0, 1, 2, 3, 4, 5]

    num_objectives: int = 3
    _ref_point = [0.40563, 0.34200, 0.75031]
    _max_hv = 0.0018603

    def __init__(
        self,
        noise_std: None | float | list[float] = None,
        negate: bool = False,
        dtype: torch.dtype = torch.double,
    ) -> None:
        r"""Data mixture curriculum optimization for Qwen3-4B-Base

        Args:
            noise_std: Standard deviation of the observation noise. If a list is
                provided, specifies separate noise standard deviations for each
                objective in a multiobjective problem.
            negate: If True, negate the function.
            dtype: The dtype that is used for the bounds of the function.
        """

        super().__init__(
            noise_std=noise_std,
            negate=negate,
            dtype=dtype,
        )

        self.model_path, self.model_config, self.y_mean, self.y_std = (
            pull_info_from_hf_hub(self.hf_repo)
        )

        self.obj_func = MLPFunction(
            self.model_config["input_dim"],
            self.model_path,
            hidden_dim=self.model_config["hidden_dim"],
            output_dim=self.model_config["output_dim"],
        )

    def _evaluate_true(self, X: torch.Tensor) -> torch.Tensor:
        r"""Evaluate the objective using the pretrained emulator.

        Args:
            X (torch.Tensor): Input tensor of shape ``(N, dim)``.

        Returns:
            torch.Tensor: Objective tensor of shape ``(N, 3)``.
        """

        validate_simplex_product(X=X, eps=1e-5)

        y_mo = self.obj_func.evaluate_true(X)
        y_mo_st = unstandardize_y(y_mo, self.y_mean, self.y_std)
        return y_mo_st

__init__(noise_std=None, negate=False, dtype=torch.double)

Data mixture curriculum optimization for Qwen3-4B-Base

Parameters:

Name Type Description Default
noise_std None | float | list[float]

Standard deviation of the observation noise. If a list is provided, specifies separate noise standard deviations for each objective in a multiobjective problem.

 None
negate bool

If True, negate the function.

 False
dtype dtype

The dtype that is used for the bounds of the function.

 double
Source code in bolt/problems/data_mixture.py 
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
def __init__(
    self,
    noise_std: None | float | list[float] = None,
    negate: bool = False,
    dtype: torch.dtype = torch.double,
) -> None:
    r"""Data mixture curriculum optimization for Qwen3-4B-Base

    Args:
        noise_std: Standard deviation of the observation noise. If a list is
            provided, specifies separate noise standard deviations for each
            objective in a multiobjective problem.
        negate: If True, negate the function.
        dtype: The dtype that is used for the bounds of the function.
    """

    super().__init__(
        noise_std=noise_std,
        negate=negate,
        dtype=dtype,
    )

    self.model_path, self.model_config, self.y_mean, self.y_std = (
        pull_info_from_hf_hub(self.hf_repo)
    )

    self.obj_func = MLPFunction(
        self.model_config["input_dim"],
        self.model_path,
        hidden_dim=self.model_config["hidden_dim"],
        output_dim=self.model_config["output_dim"],
    )

bolt.DMCurriculumHet

Data mixture selection with curriculum and heteroscedastic noise.

First set of data mixture is used to train from 0 to 5M tokens. Second set of data mixture is used to train from 5M to 10M tokens.

3 types of datasets are used
  • instruction following (IF): allenai/tulu-3-sft-personas-instruction-following
  • math: allenai/tulu-3-sft-personas-math
  • code: allenai/tulu-3-sft-personas-code
6 parameters
  1. IF proportion 1 (float, [0, 1])
  2. Math proportion 1 (float, [0, 1])
  3. Code proportion 1 (float, [0, 1])
  4. IF proportion 2 (float, [0, 1])
  5. Math proportion 2 (float, [0, 1])
  6. Code proportion 2 (float, [0, 1])

Parameters 1,2,3 must sum to 1. So must parameters 4,5,6. (Simplex)

Single objective: evaluation results on MATH-500 (minerva format)

Example usage: 

import torch
from bolt import DMCurriculumHet

prob = DMCurriculumHet()

X = torch.Tensor([[0.15, 0.2, 0.65, 0.6, 0.2, 0.2]])
y = prob(X)

Source code in bolt/problems/data_mixture.py 
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
class DMCurriculumHet(HeteroscedasticTestProblem, LLMTestProblem):
    r"""Data mixture selection with curriculum and heteroscedastic noise.

    First set of data mixture is used to train from 0 to 5M tokens.
    Second set of data mixture is used to train from 5M to 10M tokens.

    3 types of datasets are used:
        - instruction following (IF): allenai/tulu-3-sft-personas-instruction-following
        - math: allenai/tulu-3-sft-personas-math
        - code: allenai/tulu-3-sft-personas-code

    6 parameters:
        1. IF proportion 1 (float, [0, 1])
        2. Math proportion 1 (float, [0, 1])
        3. Code proportion 1 (float, [0, 1])
        4. IF proportion 2 (float, [0, 1])
        5. Math proportion 2 (float, [0, 1])
        6. Code proportion 2 (float, [0, 1])

    Parameters 1,2,3 must sum to 1. So must parameters 4,5,6. (Simplex)

    Single objective: evaluation results on MATH-500 (minerva format)

    Example usage:
    ```python
    import torch
    from bolt import DMCurriculumHet

    prob = DMCurriculumHet()

    X = torch.Tensor([[0.15, 0.2, 0.65, 0.6, 0.2, 0.2]])
    y = prob(X)
    ```

    """

    name = "dm_curriculum_heteroscedastic"
    hf_repo = "chewwt/dm_qwen4b_emulator"
    hf_repo_noise = "chewwt/dm_qwen4b_noise_emulator"

    dim = 6
    _bounds = [
        (0.0, 1.0),  # if_prop1
        (0.0, 1.0),  # math_prop1
        (0.0, 1.0),  # code_prop1
        (0.0, 1.0),  # if_prop2
        (0.0, 1.0),  # math_prop2
        (0.0, 1.0),  # code_prop2
    ]

    _check_grad_at_opt: bool = True
    continuous_inds = [0, 1, 2, 3, 4, 5]

    _optimal_value = 0.52161
    _optimizers = [(0.5311, 0.4689, 0.0000, 0.2176, 0.7824, 0.0000)]

    def __init__(
        self,
        noise_std: None | float | list[float] = None,
        negate: bool = False,
        dtype: torch.dtype = torch.double,
    ) -> None:
        r"""Data mixture curriculum optimization for Qwen3-4B-Base

        Args:
            noise_std: Standard deviation of the observation noise. Ignored
                in this problem as noise_std is output from noise model.
            negate: If True, negate the function.
            dtype: The dtype that is used for the bounds of the function.
        """

        super().__init__(
            noise_std=noise_std,
            negate=negate,
            dtype=dtype,
        )

        self.model_path, self.model_config, self.y_mean, self.y_std = (
            pull_info_from_hf_hub(self.hf_repo)
        )
        self.model_path_noise = hf_hub_download(
            self.hf_repo_noise, "noise_model.safetensors"
        )

        self.obj_func = MLPFunction(
            self.model_config["input_dim"],
            self.model_path,
            hidden_dim=self.model_config["hidden_dim"],
            output_dim=self.model_config["output_dim"],
        )

        # noise func only for math std
        self.noise_func = KRFunction(self.model_path_noise, scale_factor=0.1)

    def _evaluate_true(self, X: torch.Tensor) -> torch.Tensor:
        r"""Evaluate the objective using the pretrained emulator.

        Args:
            X (torch.Tensor): Input tensor of shape ``(N, dim)``.

        Returns:
            torch.Tensor: Objective tensor of shape ``(N, 3)``.
        """

        validate_simplex_product(X=X, eps=1e-5)

        y_mo = self.obj_func.evaluate_true(X)[:, 1]
        y_mo_st = unstandardize_y(y_mo, self.y_mean[1], self.y_std[1])
        return y_mo_st[:, None]

    def _evaluate_noise(self, X: torch.Tensor) -> torch.Tensor:
        r"""Evaluate the noise function using the pretrained noise emulator.

        Args:
            X (torch.Tensor): Input tensor of shape ``(N, dim)``.

        Returns:
            torch.Tensor: Noise std tensor of shape ``(N, 3)``.

        """

        validate_simplex_product(X=X, eps=1e-5)

        # X for noise model takes in [if_prop1, math_prop1, math_prop2] only
        # predicts math std only
        y_noise = self.noise_func.evaluate_true(X[:, [0, 1, 4]]).clamp(min=1e-6)

        return y_noise[:, None]

__init__(noise_std=None, negate=False, dtype=torch.double)

Data mixture curriculum optimization for Qwen3-4B-Base

Parameters:

Name Type Description Default
noise_std None | float | list[float]

Standard deviation of the observation noise. Ignored in this problem as noise_std is output from noise model.

 None
negate bool

If True, negate the function.

 False
dtype dtype

The dtype that is used for the bounds of the function.

 double
Source code in bolt/problems/data_mixture.py 
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
def __init__(
    self,
    noise_std: None | float | list[float] = None,
    negate: bool = False,
    dtype: torch.dtype = torch.double,
) -> None:
    r"""Data mixture curriculum optimization for Qwen3-4B-Base

    Args:
        noise_std: Standard deviation of the observation noise. Ignored
            in this problem as noise_std is output from noise model.
        negate: If True, negate the function.
        dtype: The dtype that is used for the bounds of the function.
    """

    super().__init__(
        noise_std=noise_std,
        negate=negate,
        dtype=dtype,
    )

    self.model_path, self.model_config, self.y_mean, self.y_std = (
        pull_info_from_hf_hub(self.hf_repo)
    )
    self.model_path_noise = hf_hub_download(
        self.hf_repo_noise, "noise_model.safetensors"
    )

    self.obj_func = MLPFunction(
        self.model_config["input_dim"],
        self.model_path,
        hidden_dim=self.model_config["hidden_dim"],
        output_dim=self.model_config["output_dim"],
    )

    # noise func only for math std
    self.noise_func = KRFunction(self.model_path_noise, scale_factor=0.1)