fix(collector): persist task cursor across PPO updates - was resetting to 0 each update so only first IC_ROLLOUTS tasks were ever trained on; now round-robins full shard (3 rollouts x 60 updates = 180 visits over 127-task shard = full coverage); also log first/last 5 task ids at startup so coverage is visible

colab/train_lib.py

@@ -535,6 +535,11 @@ class IncidentRolloutCollector:
     critic: LLMCritic
     tasks:  list[str]
     max_steps_per_ep: int = 16
+    # Persistent cursor — advances ACROSS update calls so we round-robin the
+    # whole task list instead of always picking tasks[0..n_episodes-1] at
+    # every PPO update. Without this, IC_ROLLOUTS=3 with 127 shard tasks would
+    # train on only the first 3 forever.
+    _cursor: int = 0
 
     def collect(self, n_episodes: int,
                 progress_cb=None) -> list[Transition]:
@@ -547,7 +552,8 @@ class IncidentRolloutCollector:
         env = IncidentCommanderEnv(use_mock=True)
         transitions: list[Transition] = []
         for ep in range(n_episodes):
-            tid = self.tasks[ep % len(self.tasks)]
+            tid = self.tasks[self._cursor % len(self.tasks)]
+            self._cursor += 1
             obs_struct = env.reset(tid)
             obs_text   = self._obs_to_text(env, obs_struct)
             for t in range(self.max_steps_per_ep):

scripts/run_training.py

@@ -110,6 +110,8 @@ if _n_shards > 1:
     print(f"[hfjob] sharded {_shard}/{_n_shards} → {len(tasks)} tasks")
 
 print(f"[hfjob] task_mode={TASK_MODE}  count={len(tasks)}")
+print(f"[hfjob] first 5 task ids: {tasks[:5]}")
+print(f"[hfjob] last  5 task ids: {tasks[-5:]}")
 
 # ── Optional warm-start from a prior phase's HF model repo ─────────────
 init_repo = os.environ.get("IC_INIT_ADAPTER_REPO", "").strip()