fix tagged_lm_evaluate (#2979)

Fixes #2976

Author: RohithKuditipudi

lib/levanter/src/levanter/eval.py

@@ -279,14 +279,22 @@ def eval_callback(step: StepInfo):
             metrics_file = os.path.join(checkpoint_path, "eval_metrics.jsonl")
             fs, _, _ = fsspec.get_fs_token_paths(metrics_file)
             fs.makedirs(checkpoint_path, exist_ok=True)
-            with fs.open(metrics_file, "a") as f:
+
+            if fs.exists(metrics_file):
+                with fs.open(metrics_file, "r") as f:
+                    content = f.read()
+            else:
+                content = ""
+
+            with fs.open(metrics_file, "w") as f:
                 # Convert numpy/jax floats to Python floats for JSON serialization
                 serializable_metrics = {
                     k: float(v) if isinstance(v, (np.floating, jnp.floating)) else v
                     for k, v in metrics_to_write.items()
                 }
                 record = {"step": int(step_count), **serializable_metrics}
-                f.write(json.dumps(record, sort_keys=True) + "\n")
+                content += json.dumps(record, sort_keys=True) + "\n"
+                f.write(content)
 
         return