config: scripts/job_train.sh \u2014 Path 3 (qwen_0_5b_smoke + GRPO max_steps=50) for sanity-check run before burning 1.5B credits

scripts/job_train.sh

@@ -36,25 +36,25 @@ python -c "import urllib.request, sys; sys.exit(0 if urllib.request.urlopen('htt
 python -m train.smoke_test --env-url http://localhost:8001
 grep -q "VERDICT: GO" outputs/smoke/SMOKE_RESULT.md || (cat outputs/smoke/SMOKE_RESULT.md; exit 1)
 
-# SFT on 1.5B
+# SFT (Path 3 — 0.5B sanity check; switch to qwen_1_5b for Path 1)
 python <<'PY'
 from hydra import initialize, compose
 from train.model_load import load_model_and_tokenizer
 from train.sft import run_sft
 with initialize(config_path="../configs", version_base="1.3"):
-    cfg = compose(config_name="config", overrides=["model=qwen_1_5b","train=sft"])
+    cfg = compose(config_name="config", overrides=["model=qwen_0_5b_smoke","train=sft"])
 m, t = load_model_and_tokenizer(cfg)
 print("SFT adapter:", run_sft(cfg, m, t))
 PY
 
-# GRPO 400 steps
+# GRPO 50 steps (Path 3 — 0.5B sanity check; switch to qwen_1_5b + max_steps=400 for Path 1)
 python <<'PY'
 from hydra import initialize, compose
 from train.model_load import load_model_and_tokenizer
 from train.grpo import run_grpo
 from rewards.compose import make_reward_fn
 with initialize(config_path="../configs", version_base="1.3"):
-    cfg = compose(config_name="config", overrides=["model=qwen_1_5b","train=grpo"])
+    cfg = compose(config_name="config", overrides=["model=qwen_0_5b_smoke","train=grpo","train.max_steps=50"])
 m, t = load_model_and_tokenizer(cfg)
 print("GRPO merged:", run_grpo(cfg, m, t, make_reward_fn(cfg.reward), "http://localhost:8001"))
 PY