Merge pull request #201 from macrocosm-os/dev

Release 4.6.1
macrocosm-os · Nov 13, 2024 · 49138f9 · 49138f9
2 parents 69458c0 + b9901e8
commit 49138f9
Show file tree

Hide file tree

Showing 2 changed files with 11 additions and 6 deletions.
diff --git a/constants/__init__.py b/constants/__init__.py
@@ -34,7 +34,7 @@
 # ---------------------------------
 
 # Release
-__version__ = "4.6.0"
+__version__ = "4.6.1"
 
 # Validator schema version
 __validator_version__ = "4.6.0"

diff --git a/neurons/validator.py b/neurons/validator.py
@@ -878,7 +878,7 @@ async def run_step(self):
         # Use sample packing.
         pack_samples = True
         pages_per_eval = constants.pages_per_eval_pack
-        
+
         # Try to synchronize the data used by validators.
         try:
             seed = metagraph_utils.get_hash_of_sync_block(
@@ -925,7 +925,7 @@ async def run_step(self):
             )
             bt.logging.debug(f"14b* Batch size is {len(batches_14b_star[0])}")
 
-            #This is useful for logging to wandb
+            # This is useful for logging to wandb
             pages_14b_star = dataloader_14b_star.get_page_names()
 
             bt.logging.debug(f"14b* Pages used are {pages_14b_star}")
@@ -1057,6 +1057,11 @@ async def run_step(self):
                     competition=competition_14b_star,
                 )
             )
+            # Since we choose the models to keep based on weight only if they are in the win_rate dict,
+            # We need to add anything only winning in 14B* to it. We use a wr of 0 so it is only based on weight.
+            win_rate.update(
+                {key: 0.0 for key in win_rate_14b_star if key not in win_rate}
+            )
 
         # Get ids for all competitions in the schedule.
         active_competition_ids = set([comp.id for comp in competition_schedule])
@@ -1333,9 +1338,9 @@ def log_step(
             for dataset_name, avg_loss in (
                 uid_to_state[uid].avg_loss_per_dataset().items()
             ):
-                step_log["uid_data"][str(uid)]["dataset_perf"][
-                    f"{dataset_name}"
-                ] = {'average_loss': avg_loss}
+                step_log["uid_data"][str(uid)]["dataset_perf"][f"{dataset_name}"] = {
+                    "average_loss": avg_loss
+                }
         table = Table(title="Step", expand=True)
         table.add_column("uid", justify="right", style="cyan", no_wrap=True)
         table.add_column("hf", style="magenta", overflow="fold")