_runs/launch_baseline_push_v6.sh

#!/usr/bin/env bash
# Wave-6: push baseline 1.5B past solve=0.40 by porting the latent's winning
# reward shaping + multi-value oversampling into the vanilla baseline pipeline.
#
# Diagnosis from v4/v5 logs:
#   At plateau, eval shows `avg_set_size=1.000` for every step. The model
#   is predicting only ONE value per cell even when the target is multi-valued.
#   Per-cell exact pinned at 0.95 → solve = 0.95^20 ≈ 0.36.
#   Same failure mode the latent's `s3_grpo_sharp_rwd` recipe fixed:
#     exact_match_bonus + cardinality_mismatch_penalty + penalty_missing
#   plus SFT-side multi_value_oversample_factor=5 (and target_size_min=2 for
#   the most aggressive variant).
#
# 8 variants on GPUs 0..7. All seed from existing v4 best ckpts so we don't
# burn cycles redoing S2.
set -euo pipefail

ROOT=/home/ubuntu/curriculum_cot
SWEEP_ROOT=$ROOT/_runs/baseline_1p5b_v4_20260523_184952
PIPELINE=$ROOT/_runs/baseline_1p5b_pipeline_v4.sh

# --- v4 anchors ----
PIPE_V_S3SFT_LATEST=$SWEEP_ROOT/pipe_v_sft_extend/s3_sft/checkpoint-step-04000
PIPE_M_S3SFT_LATEST=$SWEEP_ROOT/pipe_m_s3sft_from_b/s3_sft/checkpoint-step-02400
PIPE_V_S3GRPO_BEST=$SWEEP_ROOT/pipe_v_sft_extend/s3_grpo/checkpoint-1000   # step 1050 was 0.40 peak; 1000 is closest saved
PIPE_M_S3GRPO_BEST=$SWEEP_ROOT/pipe_m_s3sft_from_b/s3_grpo/checkpoint-200  # peak per pipe_m logs
PIPE_O_S3SFT_LATEST=$SWEEP_ROOT/pipe_o_s3sft_lr5e6/s3_sft/checkpoint-step-02400
CKPT_LR5E5=$ROOT/checkpoints/sudoku-9x9-20empty-baseline-1p5b-sweep/baseline_lr5e5_lowsft_v3/s2_sft_v3/checkpoint-step-03000

for c in "$PIPE_V_S3SFT_LATEST" "$PIPE_M_S3SFT_LATEST" "$PIPE_V_S3GRPO_BEST" "$PIPE_M_S3GRPO_BEST" "$PIPE_O_S3SFT_LATEST"; do
  [[ -d "$c" ]] || { echo "MISSING: $c"; exit 1; }
done

launch() {
  local gpu="$1" variant="$2"; shift 2
  local out=$SWEEP_ROOT/$variant; mkdir -p "$out"
  nohup env ROOT="$ROOT" VARIANT="$variant" GPU="$gpu" S2_SFT_CKPT="$CKPT_LR5E5" \
    OUTPUT_ROOT="$out" USE_WANDB=0 WANDB_MODE=offline "$@" \
    bash "$PIPELINE" </dev/null >"$out/nohup.log" 2>&1 &
  local pid=$!
  echo "$pid $gpu $variant" >> "$SWEEP_ROOT/PIDS.txt"
  disown $pid 2>/dev/null || true
  printf 'GPU %s -> %s pid=%s\n' "$gpu" "$variant" "$pid"
}

# === GRPO continuations (the high-leverage knob) ===

# v6_a (GPU 0): continue best v4 GRPO with the FULL latent recipe.
# card_pen=1.0 + missing=0.75 + exact_b=2.0; LR slightly lower than v4 to be safe.
launch 0 v6_a_grpo_v_card \
  START_PHASE=s3_grpo S3_GRPO_INIT="$PIPE_V_S3GRPO_BEST" \
  GRPO_LR=2e-6 GRPO_BS=32 GRPO_GA=1 GRPO_NG=8 \
  PENALTY_MISSING=0.75 EXACT_MATCH_BONUS=2.0 CARD_MISMATCH_PEN=1.0 \
  S3_GRPO_MAX_STEPS=2000

# v6_b (GPU 1): "sharp" version — mirror s3_grpo_sharp_rwd's stronger weights.
launch 1 v6_b_grpo_v_sharp \
  START_PHASE=s3_grpo S3_GRPO_INIT="$PIPE_V_S3GRPO_BEST" \
  GRPO_LR=2e-6 GRPO_BS=32 GRPO_GA=1 GRPO_NG=8 \
  PENALTY_MISSING=1.0 EXACT_MATCH_BONUS=4.0 CARD_MISMATCH_PEN=3.0 \
  S3_GRPO_MAX_STEPS=2000

# v6_c (GPU 2): full recipe but from pipe_v's S3 SFT (fresh GRPO, not continuation).
launch 2 v6_c_grpo_vsft_card \
  START_PHASE=s3_grpo S3_GRPO_INIT="$PIPE_V_S3SFT_LATEST" \
  GRPO_LR=5e-6 GRPO_BS=32 GRPO_GA=1 GRPO_NG=8 \
  PENALTY_MISSING=0.75 EXACT_MATCH_BONUS=2.0 CARD_MISMATCH_PEN=1.0 \
  S3_GRPO_MAX_STEPS=2000

# v6_d (GPU 3): same recipe but from pipe_m's S3 SFT (different lineage; champion).
launch 3 v6_d_grpo_msft_card \
  START_PHASE=s3_grpo S3_GRPO_INIT="$PIPE_M_S3SFT_LATEST" \
  GRPO_LR=5e-6 GRPO_BS=32 GRPO_GA=1 GRPO_NG=8 \
  PENALTY_MISSING=0.75 EXACT_MATCH_BONUS=2.0 CARD_MISMATCH_PEN=1.0 \
  S3_GRPO_MAX_STEPS=2000

# === SFT push w/ oversample (the data-side knob) ===

# v6_e (GPU 4): continue pipe_v S3 SFT with oversample=5. Mirrors r1_sft_c_oversample5.
launch 4 v6_e_sft_v_oversample5 \
  START_PHASE=s3_sft S3_SFT_INIT="$PIPE_V_S3SFT_LATEST" \
  SFT_LR_S3=2e-6 SFT_BS=16 SFT_GA=1 \
  SFT_OVERSAMPLE=5 \
  S3_SFT_MAX_STEPS=2500 \
  GRPO_LR=2e-6 GRPO_BS=32 GRPO_GA=1 GRPO_NG=8 \
  PENALTY_MISSING=0.75 EXACT_MATCH_BONUS=2.0 CARD_MISMATCH_PEN=1.0 \
  S3_GRPO_MAX_STEPS=1500

# v6_f (GPU 5): same but oversample=8 (more aggressive).
launch 5 v6_f_sft_v_oversample8 \
  START_PHASE=s3_sft S3_SFT_INIT="$PIPE_V_S3SFT_LATEST" \
  SFT_LR_S3=2e-6 SFT_BS=16 SFT_GA=1 \
  SFT_OVERSAMPLE=8 \
  S3_SFT_MAX_STEPS=2500 \
  GRPO_LR=2e-6 GRPO_BS=32 GRPO_GA=1 GRPO_NG=8 \
  PENALTY_MISSING=0.75 EXACT_MATCH_BONUS=2.0 CARD_MISMATCH_PEN=1.0 \
  S3_GRPO_MAX_STEPS=1500

# v6_g (GPU 6): oversample=5 + train_target_size_min=2 (only multi-value cells).
# This is the most surgical variant — focus all training mass on the failing cells.
launch 6 v6_g_sft_v_mv_only \
  START_PHASE=s3_sft S3_SFT_INIT="$PIPE_V_S3SFT_LATEST" \
  SFT_LR_S3=1e-6 SFT_BS=16 SFT_GA=1 \
  SFT_OVERSAMPLE=5 SFT_TGT_MIN=2 \
  S3_SFT_MAX_STEPS=2000 \
  GRPO_LR=2e-6 GRPO_BS=32 GRPO_GA=1 GRPO_NG=8 \
  PENALTY_MISSING=0.75 EXACT_MATCH_BONUS=2.0 CARD_MISMATCH_PEN=1.0 \
  S3_GRPO_MAX_STEPS=1500

# v6_h (GPU 7): same as v6_a but with even more steps + KL anchor for stability.
# The latent best (s3_grpo_baseline) ran with beta=0.0; we know KL>0 hurts long term.
# But here we want to see whether the new shape rewards survive more steps without
# regression. Use a small beta (0.01) for gentle anchoring.
launch 7 v6_h_grpo_v_card_long \
  START_PHASE=s3_grpo S3_GRPO_INIT="$PIPE_V_S3GRPO_BEST" \
  GRPO_LR=2e-6 GRPO_BS=32 GRPO_GA=1 GRPO_NG=8 \
  GRPO_BETA=0.01 \
  PENALTY_MISSING=0.75 EXACT_MATCH_BONUS=2.0 CARD_MISMATCH_PEN=1.0 \
  S3_GRPO_MAX_STEPS=3000

echo
echo "=== launched ==="
tail -8 "$SWEEP_ROOT/PIDS.txt"