Upload runpod/create-template.sh with huggingface_hub

runpod/create-template.sh

@@ -313,12 +313,13 @@ elif profile in (
         "__tool_context_mix__"
     )
     if tool_routing_rl:
-        # Answer meal v10: init v9 pulse-250; OpenRouter judge critical; adapted traces.
-        steps, warmup, wall, label, lr = "300", "20", "10800", "v10-tool-answer-rl", "5e-6"
+        # Answer meal v11: init v10 final; up-weight full-loop CE; gold-answer on gate fail;
+        # 1000 steps or early-stop on consistent E.
+        steps, warmup, wall, label, lr = "1000", "40", "43200", "v11-tool-answer-rl", "5e-6"
         init_file = (
-            "training-runs/sft-chat-v9-tool-routing-rl-8b-20260813-123225/model_pulse_250.pt"
+            "training-runs/sft-chat-v10-tool-answer-rl-8b-20260813-183513/model_final.pt"
         )
-        pulses = "100,200"
+        pulses = "200,400,600,800"
         datasets = "__tool_routing_rl_mix__"
         weights = "1.0"
         min_lr = "1e-6"
@@ -637,11 +638,11 @@ elif profile in (
         env["REJECT_TOOL_TRACES"] = "0"
         env["ALLOW_PHANTOM_TOOLS"] = "1"
     if tool_routing_rl:
-        # routing-rl: Call-RL free-gen primary; snack CE tiny.
+        # routing-rl: Call-RL free-gen primary; stronger loop CE for complete answers (v11).
         # batch stays 4 — batch=6 + max_new=256 OOMed/SIGABRT (rc=134) on PRO 6000.
         env["EMBED_RL_MAX_NEW"] = "256"
         env["EMBED_RL_BATCH"] = "4"
-        env["BEHAVIOR_SFT_SCALE"] = "0.02"
+        env["BEHAVIOR_SFT_SCALE"] = "0.35"
         env["PUSH"] = "1"
         env["ASYNC_BUCKET_UPLOAD"] = "1"
         env["OPENROUTER_JUDGE_TIMEOUT"] = "15"
@@ -678,12 +679,12 @@ elif profile in (
         )
     elif tool_routing_rl:
         readme = (
-            "# Scugnizz Llama-PCS chat tool-answer-rl v10 (judge critical)\n\n"
-            "Init from **v9 pulse-250**. Mix: 50% adapt call / 40% adapt loops / 10% schema "
-            "(real Hermes payloads — no routing synth). Call-RL ladder A–E: MiniLM "
-            "prefilter (answer↔tool + answer↔Q) then OpenRouter judge; timeout 15s → D+. "
-            "**300** steps, pulses **100,200**, LR **5e-6**→**1e-6**. "
-            "Requires OPENROUTER_API_KEY env inject.\n"
+            "# Scugnizz Llama-PCS chat tool-answer-rl v11 (complete examples + judge)\n\n"
+            "Init from **v10 final**. Mix ~30% adapt call / **60% adapt loops** / 10% schema. "
+            "BEHAVIOR_SFT_SCALE **0.35**. Call-RL: on gate-fail after right call, CE Hermes "
+            "**gold_answer** (complete traj). Same OpenRouter judge ladder; early-stop if "
+            "consistent E (10 straight steps with crit + 20/40 window). "
+            "**1000** steps (or early E), pulses **200,400,600,800**. Keep **4096** ctx.\n"
         )
     elif tool_hermes_harness:
         readme = (