wandb does not work with verifiers integration

[tmabraham]

Running python -m tinker_cookbook.recipes.verifiers_rl.train errors out if wandb logging is enabled:

Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/home/tanishq/tinker-cookbook/tinker_cookbook/recipes/verifiers_rl/train.py", line 184, in <module>
    asyncio.run(cli_main(cli_config, None))
  File "/home/tanishq/.local/share/uv/python/cpython-3.12.11-linux-x86_64-gnu/lib/python3.12/asyncio/runners.py", line 195, in run
    return runner.run(main)
           ^^^^^^^^^^^^^^^^
  File "/home/tanishq/.local/share/uv/python/cpython-3.12.11-linux-x86_64-gnu/lib/python3.12/asyncio/runners.py", line 118, in run
    return self._loop.run_until_complete(task)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/tanishq/.local/share/uv/python/cpython-3.12.11-linux-x86_64-gnu/lib/python3.12/asyncio/base_events.py", line 691, in run_until_complete
    return future.result()
           ^^^^^^^^^^^^^^^
  File "/home/tanishq/tinker-cookbook/tinker_cookbook/recipes/verifiers_rl/train.py", line 179, in cli_main
    await train.main(cfg)
  File "/home/tanishq/tinker-cookbook/tinker_cookbook/utils/trace.py", line 332, in async_wrapper
    return await func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/tanishq/tinker-cookbook/tinker_cookbook/rl/train.py", line 990, in main
    ml_logger = ml_log.setup_logging(
                ^^^^^^^^^^^^^^^^^^^^^
  File "/home/tanishq/tinker-cookbook/tinker_cookbook/utils/ml_log.py", line 470, in setup_logging
    ml_logger.log_hparams(config)
  File "/home/tanishq/tinker-cookbook/tinker_cookbook/utils/ml_log.py", line 349, in log_hparams
    logger.log_hparams(config)
  File "/home/tanishq/tinker-cookbook/tinker_cookbook/utils/ml_log.py", line 230, in log_hparams
    wandb.config.update(dump_config(config))
  File "/home/tanishq/tinker-cookbook/.venv/lib/python3.12/site-packages/wandb/sdk/wandb_config.py", line 187, in update
    sanitized = self._update(d, allow_val_change)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/tanishq/tinker-cookbook/.venv/lib/python3.12/site-packages/wandb/sdk/wandb_config.py", line 180, in _update
    sanitized = self._sanitize_dict(
                ^^^^^^^^^^^^^^^^^^^^
  File "/home/tanishq/tinker-cookbook/.venv/lib/python3.12/site-packages/wandb/sdk/wandb_config.py", line 267, in _sanitize_dict
    k, v = self._sanitize(k, v, allow_val_change)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/tanishq/tinker-cookbook/.venv/lib/python3.12/site-packages/wandb/sdk/wandb_config.py", line 288, in _sanitize
    raise config_util.ConfigError(
wandb.sdk.lib.config_util.ConfigError: Attempted to change value of key "dataset_builder" from {'vf_env': '<verifiers.envs.singleturn_env.SingleTurnEnv object at 0x7fd46752da30>',
'groups_per_batch': 128, 'dataset_n': -1, 'dataset_seed': None} to {'vf_env': '<verifiers.envs.singleturn_env.SingleTurnEnv object at 0x7fd4677eb800>', 'groups_per_batch': 128, 'dataset_n': -1,
 'dataset_seed': None}
If you really want to do this, pass allow_val_change=True to config.update()

It seems the reason seems to be VerifiersRLDatasetBuilder is a chz.chz class and all the fields get saved to wandb but this includes vf_env the full verifiers env... i couldn't find any good solution to avoid this and came up with some hacky solution with Claude Code's help:

-@chz.chz
 class VerifiersRLDatasetBuilder(RLDatasetBuilder):
-    vf_env: vf.Environment
-    groups_per_batch: int
-    dataset_n: int
-    dataset_seed: int | None
+    def __init__(
+        self,
+        vf_env: vf.Environment,
+        groups_per_batch: int,
+        dataset_n: int,
+        dataset_seed: int | None,
+    ):
+        # Use object.__setattr__ to bypass frozen instance restrictions from parent chz class
+        object.__setattr__(self, 'vf_env', vf_env)
+        object.__setattr__(self, 'groups_per_batch', groups_per_batch)
+        object.__setattr__(self, 'dataset_n', dataset_n)
+        object.__setattr__(self, 'dataset_seed', dataset_seed)
+
+    def to_dict(self):
+        """Return a dictionary representation excluding the vf_env object."""
+        return {
+            "groups_per_batch": self.groups_per_batch,
+            "dataset_n": self.dataset_n,
+            "dataset_seed": self.dataset_seed,
+        }

Let me know if there may be better approaches for addressing this issue otherwise I'm happy to PR this change directly as well.

cc: @willccbb