Sync contents

Author: Tiiiger

.sync_state

@@ -1,4 +1,4 @@
 {
-  "last_synced_sha": "e039f707084c266fc64d92b2228da6f0a30710b9",
-  "last_sync_time": "2025-10-02T01:33:33.157309"
+  "last_synced_sha": "09d45a371c67ee32d2a31ffcdc6709a9b8e10a18",
+  "last_sync_time": "2025-10-05T23:54:36.006757"
 }

tinker_cookbook/hyperparam_utils.py

@@ -148,9 +148,9 @@ def get_lr(model_name: str, is_lora: bool = True) -> float:
     lora_multiplier = 10.0
 
     lr = base_lr * lora_multiplier if is_lora else base_lr
-    if "llama" in model_name:
+    if "llama" in model_name.lower():
         exponent_model = 0.781
-    elif "qwen" in model_name:
+    elif "qwen" in model_name.lower():
         exponent_model = 0.0775
     else:
         assert False, f"Unknown model: {model_name}"

tinker_cookbook/recipes/math_rl/arithmetic_env.py

@@ -1,4 +1,5 @@
 from functools import partial
+from typing import Sequence
 
 import chz
 import numpy as np
@@ -62,7 +63,7 @@ def __init__(
         self.n_batches = n_batches
         self.include_fewshot = include_fewshot
 
-    def get_batch(self, index: int) -> list[EnvGroupBuilder]:
+    def get_batch(self, index: int) -> Sequence[EnvGroupBuilder]:
         self._rng.seed(index)
         return [self._make_env_group_builder(self._rng) for _ in range(self.batch_size)]
 

tinker_cookbook/recipes/math_rl/math_env.py

@@ -1,7 +1,7 @@
 import math
 import re
 from functools import partial
-from typing import Literal, cast
+from typing import Literal, Sequence, cast
 
 import chz
 from datasets import Dataset, concatenate_datasets, get_dataset_config_names, load_dataset
@@ -153,7 +153,7 @@ def __init__(
         self.renderer = renderer
         self.convo_prefix = convo_prefix
 
-    def get_batch(self, index: int) -> list[EnvGroupBuilder]:
+    def get_batch(self, index: int) -> Sequence[EnvGroupBuilder]:
         batch_start = index * self.batch_size
         batch_end = min((index + 1) * self.batch_size, len(self.ds))
         assert batch_start < batch_end, "Incorrect batch size"
@@ -329,7 +329,7 @@ def __init__(
     def question_suffix(cls) -> str:
         return " Provide a numerical answer without units, written inside \\boxed{}."
 
-    def get_batch(self, index: int) -> list[EnvGroupBuilder]:
+    def get_batch(self, index: int) -> Sequence[EnvGroupBuilder]:
         batch_start = index * self.batch_size
         batch_end = min((index + 1) * self.batch_size, len(self.ds))
         assert batch_start < batch_end, "Incorrect batch size"

tinker_cookbook/recipes/multiplayer_rl/guess_number/env.py

@@ -112,7 +112,7 @@ class GuessNumberDataset(RLDataset):
     batch_size: int
     group_size: int
 
-    def get_batch(self, index: int) -> list[EnvGroupBuilder]:
+    def get_batch(self, index: int) -> Sequence[EnvGroupBuilder]:
         return [
             GuessNumberEnvGroupBuilder(
                 answer=self.answers[index * self.batch_size + i],

tinker_cookbook/recipes/multiplayer_rl/text_arena/env.py

@@ -234,7 +234,7 @@ def __init__(self, batch_size: int, builder: TwoPlayerEnvGroupBuilder, num_datap
             "num_datapoints must be divisible by num_players"
         )
 
-    def get_batch(self, index: int) -> list[EnvGroupBuilder]:
+    def get_batch(self, index: int) -> Sequence[EnvGroupBuilder]:
         return [
             self.builder
             for i in range(self.batch_size // self.builder.num_players)

tinker_cookbook/recipes/multiplayer_rl/twenty_questions/env.py

@@ -165,7 +165,7 @@ class TwentyQuestionsDataset(RLDataset):
     batch_size: int
     group_size: int
 
-    def get_batch(self, index: int) -> list[EnvGroupBuilder]:
+    def get_batch(self, index: int) -> Sequence[EnvGroupBuilder]:
         return [
             TwentyQuestionsEnvGroupBuilder(
                 answerer=self.answerer,

tinker_cookbook/recipes/tool_use/search/search_env.py

@@ -6,7 +6,7 @@
 import string
 from functools import partial, reduce
 from pathlib import Path
-from typing import Literal, TypedDict, cast
+from typing import Literal, Sequence, TypedDict, cast
 
 import chz
 import pandas as pd
@@ -315,7 +315,7 @@ def __init__(
         rng = random.Random(self.seed)
         rng.shuffle(self.ds)
 
-    def get_batch(self, index: int) -> list[EnvGroupBuilder]:
+    def get_batch(self, index: int) -> Sequence[EnvGroupBuilder]:
         return [
             self._make_env_group_builder(row, self.group_size)
             for row in self.ds[index * self.batch_size : (index + 1) * self.batch_size]

tinker_cookbook/rl/train.py

@@ -8,7 +8,7 @@
 import logging
 import os
 import time
-from typing import Any, Callable, List
+from typing import Any, Callable, List, Sequence
 
 import chz
 import numpy as np
@@ -567,7 +567,7 @@ async def save_checkpoint_and_get_sampling_client(
 
 async def prepare_minibatch(
     cfg: Config,
-    env_group_builders_P: list[EnvGroupBuilder],
+    env_group_builders_P: Sequence[EnvGroupBuilder],
     trajectory_groups_P: list[TrajectoryGroup],
     tokenizer: Tokenizer,
     service_client: tinker.ServiceClient,
@@ -744,7 +744,7 @@ async def do_train_step_and_get_sampling_client(
     training_client: tinker.TrainingClient,
     service_client: tinker.ServiceClient,
     tokenizer: Tokenizer,
-    env_group_builders_P: list[EnvGroupBuilder],
+    env_group_builders_P: Sequence[EnvGroupBuilder],
     trajectory_groups_P: list[TrajectoryGroup],
 ) -> tuple[tinker.SamplingClient, dict[str, Any]]:
     metrics = {}

tinker_cookbook/rl/types.py

@@ -137,7 +137,7 @@ class RLDataset(ABC):
     """
 
     @abstractmethod
-    def get_batch(self, index: int) -> list[EnvGroupBuilder]:
+    def get_batch(self, index: int) -> Sequence[EnvGroupBuilder]:
         pass
 
     @abstractmethod