Upload folder using huggingface_hub

code/train_value_iteration.py

@@ -186,20 +186,25 @@ def main():
                    help="States are sampled from walks of 1..this many moves. Past "
                         "~20 every state is fully mixed (God's Number), so a larger "
                         "value buys nothing.")
-    p.add_argument("--update-threshold", type=float, default=0.01,
-                   help="Refresh the target network once training loss falls below "
-                        "this. This is meant to be the binding constraint: it lets "
-                        "an easy early target move on fast and gives a harder later "
-                        "one as long as it needs. Set it above the loss the net "
-                        "actually reaches and it never binds, leaving --min-update-"
-                        "steps to pin the cadence -- measured at 0.05, the net sat "
-                        "at 0.0004 and idled ~800 of every 1000 steps.")
-    p.add_argument("--min-update-steps", type=int, default=200,
+    p.add_argument("--update-threshold", type=float, default=0.15, metavar="MOVES",
+                   help="Refresh the target network once the net fits its current "
+                        "target to within this mean absolute error, IN MOVES. "
+                        "Deliberately not an MSE threshold: target values grow from "
+                        "~1 early to ~20 at full distance and MSE scales with their "
+                        "square, so the achievable loss floor rises through the run "
+                        "and no fixed MSE value works at both ends (measured: 0.0006 "
+                        "at step 400, 0.02-0.03 by step 1500). MAE in moves is "
+                        "interpretable and roughly scale-stable.")
+    p.add_argument("--min-update-steps", type=int, default=100,
                    help="Floor on the refresh interval, so one lucky batch cannot "
-                        "ratchet the target forward. Each refresh propagates "
-                        "accuracy roughly one move further from solved, so this "
-                        "floor times ~30 is the run length; measured re-convergence "
-                        "after a refresh is ~200 steps.")
+                        "ratchet the target forward.")
+    p.add_argument("--max-update-steps", type=int, default=400,
+                   help="Backstop: refresh after this many steps even if the "
+                        "threshold is unmet. Each refresh propagates accuracy about "
+                        "one move further from solved, so a threshold that stops "
+                        "being reachable stalls the frontier outright while training "
+                        "still looks healthy -- observed at 1300 steps with no "
+                        "refresh. This bounds the run at ~max*30 steps.")
     p.add_argument("--eval-every", type=int, default=2000)
     p.add_argument("--eval-n", type=int, default=200)
     p.add_argument("--eval-depths", default="3,6,10,15,20")
@@ -273,15 +278,21 @@ def main():
         opt.step()
 
         l = loss.item()
-        if l < args.update_threshold and step - last_update >= args.min_update_steps:
+        with torch.no_grad():
+            mae = (pred.float() - y).abs().mean().item()
+        since = step - last_update
+        ready = mae < args.update_threshold and since >= args.min_update_steps
+        forced = since >= args.max_update_steps
+        if ready or forced:
             target_net.load_state_dict(net.state_dict())
             updates += 1
             last_update = step
-            print(f"  step {step}: target net updated (#{updates}, loss {l:.4f})",
+            print(f"  step {step}: target net updated (#{updates}, "
+                  f"mae {mae:.3f} moves{', forced' if forced and not ready else ''})",
                   flush=True)
 
         if step % 100 == 0:
-            print(f"step {step}/{args.max_steps} loss {l:.4f} "
+            print(f"step {step}/{args.max_steps} loss {l:.4f} mae {mae:.3f} "
                   f"tgt_updates {updates} {step/(time.time()-t0):.1f} it/s", flush=True)
 
         if step % args.eval_every == 0 or step == args.max_steps:

code/vast_bootstrap.sh

@@ -105,8 +105,9 @@ PYVI
     --hidden "${HIDDEN:-512}" --layers "${LAYERS:-8}" --heads 8 \
     --batch-size "${BATCH:-1024}" --lr "${LR:-3e-4}" --max-steps "${STEPS:-30000}" \
     --scramble-depth "${WALK_DEPTH:-30}" \
-    --update-threshold "${UPDATE_THRESHOLD:-0.05}" \
-    --min-update-steps "${MIN_UPDATE_STEPS:-1000}" \
+    --update-threshold "${UPDATE_THRESHOLD:-0.15}" \
+    --min-update-steps "${MIN_UPDATE_STEPS:-100}" \
+    --max-update-steps "${MAX_UPDATE_STEPS:-400}" \
     --eval-every "${EVAL_EVERY:-1000}" --eval-n "${EVAL_N:-200}" \
     --out checkpoints/value_iter --hub-repo "${HUB_REPO:-}"
   echo "=== done ==="