Reverendo commited on
Commit
0edf707
·
verified ·
1 Parent(s): 8702f99

Upload runpod/create-template.sh with huggingface_hub

Browse files
Files changed (1) hide show
  1. runpod/create-template.sh +9 -10
runpod/create-template.sh CHANGED
@@ -313,9 +313,9 @@ elif profile in (
313
  "__tool_context_mix__"
314
  )
315
  if tool_routing_rl:
316
- # Routing RL v8: from v7 final; schema/args/trajectory Call-RL + full-loop CE.
317
- steps, warmup, wall, label, lr = "500", "25", "10800", "v8-tool-routing-rl", "5e-6"
318
- init_file = "training-runs/sft-chat-v7-tool-routing-rl-8b-20260812-184200/model_final.pt"
319
  pulses = "250"
320
  datasets = "__tool_routing_rl_mix__"
321
  weights = "1.0"
@@ -666,13 +666,12 @@ elif profile in (
666
  )
667
  elif tool_routing_rl:
668
  readme = (
669
- "# Scugnizz Llama-PCS chat tool-routing-rl v8 (Call-RL gen-first)\n\n"
670
- "Init from **v7 final**. **Gen primary:** CALL_RL every step, batch **4** "
671
- "(6 OOMed), BEHAVIOR_SFT_SCALE **0.02**. Mix: 80% call-only (40% routing / "
672
- "25% filtered adapt / 15% schema) + **20%** thin full-loop CE glue. "
673
- "F0 schema-illegal hardest; multi-tool OK if grounded; junk exact "
674
- "`The repository` / `# hello`; post-tool gen required; critical if embed≥τ. "
675
- "Vault probe every 30 steps (log only). **500** steps, pulse **250**.\n"
676
  )
677
  elif tool_hermes_harness:
678
  readme = (
 
313
  "__tool_context_mix__"
314
  )
315
  if tool_routing_rl:
316
+ # Routing RL v9: from v8 final; kill vault_search-default.
317
+ steps, warmup, wall, label, lr = "500", "25", "10800", "v9-tool-routing-rl", "5e-6"
318
+ init_file = "training-runs/sft-chat-v8-tool-routing-rl-8b-20260813-094328/model_final.pt"
319
  pulses = "250"
320
  datasets = "__tool_routing_rl_mix__"
321
  weights = "1.0"
 
666
  )
667
  elif tool_routing_rl:
668
  readme = (
669
+ "# Scugnizz Llama-PCS chat tool-routing-rl v9 (Call-RL routing tighten)\n\n"
670
+ "Init from **v8 final**. Kill **vault_search-as-default**: info Q + vault "
671
+ "is category_mismatch; named-path read vs search vs write is "
672
+ "`wrong_vault_verb`; junk `Find a ` queries. Mix: 50% routing synth / "
673
+ "15% adapt call / 10% schema / 10% routing loops / 15% adapt loops. "
674
+ "Defensive prose gate (no `</tool_call>` criticals). **500** steps, pulse **250**.\n"
 
675
  )
676
  elif tool_hermes_harness:
677
  readme = (