fix smoke.

justfile

@@ -2,8 +2,12 @@ set shell := ["bash", "-cu"]
 
 # Three seeds for headline arms; one seed for ablations.
 SEEDS_3 := "41 43 44"
-MODEL := "Qwen/Qwen2.5-1.5B"  # placeholder for tiny-dev; real runs use Qwen3.5-2B
-TINY_MODEL := "hf-internal-testing/tiny-random-LlamaForCausalLM"  # tiny+stable; swap for real arch if needed
+# H4 main: Qwen3.5-2B; if H4 falsified (vanilla hack<30%), switch to Qwen/Qwen3-4B per spec.md.
+MODEL := "Qwen/Qwen3.5-2B"
+# Compute-fit override for 96GB single-GPU (see docs/grpo_hyperparams.md §Our deviations).
+NUM_GEN := "8"
+BATCH := "16"
+TINY_MODEL := "llamafactory/tiny-random-qwen3"  # qwen3 arch, ~6M params, smoke only
 BASE := "uv run python -m projected_grpo.run"
 
 default:
@@ -47,22 +51,28 @@ queue:
     # just queue-projected-m32        # H2 sweep
 
 # Vanilla GRPO baseline, 3 seeds. H: hack rate >30% at step 200 per spec H4.
+# Real run goes through Ariahw's verl pipeline (NOT our smoke run.py).
 queue-vanilla:
     #!/usr/bin/env bash
     set -x
     for seed in {{ SEEDS_3 }}; do
-        pueue add -w "$PWD" -o 5 \
-          -l "why: H4 sanity, does Qwen3.5-2B reward-hack at all; resolve: if <30% hack rate, swap to Qwen3-4B" \
-          -- {{ BASE }} --arm=vanilla --seed=$seed --model={{ MODEL }} --steps=200
+        pueue add -w "$PWD/external/rl-rewardhacking" -o 5 \
+          -l "why: H4 sanity, does {{ MODEL }} reward-hack at all; resolve: if <30% hack rate at step 200, swap MODEL to Qwen/Qwen3-4B + reduce NUM_GEN to 4" \
+          -- uv run python scripts/run_rl_training.py no_intervention \
+              --model_id={{ MODEL }} --seed=$seed \
+              --num_generations={{ NUM_GEN }} --per_device_batch_size={{ BATCH }}
     done
 
 # Projected gradient, m=16, 3 seeds. H1 main result.
+# TODO: integrate project_grad_per_row into verl's GRPO trainer. Currently the
+# justfile recipe still calls our smoke run.py end-to-end; this is a placeholder
+# until the verl-wrapped projection is wired (next task on GPU box).
 queue-projected-m16:
     #!/usr/bin/env bash
     set -x
     for seed in {{ SEEDS_3 }}; do
         pueue add -w "$PWD" -o 4 \
-          -l "why: H1 main, gradient proj reduces hack rate >=30pp at matched pass; resolve: publish if H1 holds" \
+          -l "why: H1 main, gradient proj reduces hack rate >=30pp at matched pass; resolve: publish if H1 holds; BLOCKED: needs verl integration" \
           -- {{ BASE }} --arm=projected --m=16 --seed=$seed --model={{ MODEL }} --steps=200
     done