Add training/train_4b_v43_local.sh

training/train_4b_v43_local.sh

@@ -0,0 +1,64 @@
+#!/bin/bash
+# train_4b_v43_local.sh — mlx_lm.lora SFT for Prism Coder 4B v43 on Apple Silicon
+# Prereq: python3 build_4b_v43_corpus.py  (creates /tmp/4b_v43_data/train.jsonl + valid.jsonl)
+# v43.2 patch: resumed from 1400-iter checkpoint with surgical corpus patch
+#   (+2075 rows: param-precise, multi-turn chain, abstain, disambiguation)
+set -euo pipefail
+
+MODEL="Qwen/Qwen3-4B"
+DATA_DIR="/tmp/4b_v43_data"
+ADAPTER_DIR="/tmp/4b_v43_adapter"
+LOG="/tmp/4b_v43_train.log"
+
+# ── Sanity checks ────────────────────────────────────────────────────────────
+if [ ! -f "$DATA_DIR/train.jsonl" ]; then
+    echo "ERROR: $DATA_DIR/train.jsonl not found."
+    echo "Run:   python3 build_4b_v43_corpus.py"
+    exit 1
+fi
+
+TRAIN_ROWS=$(wc -l < "$DATA_DIR/train.jsonl")
+VALID_ROWS=$(wc -l < "$DATA_DIR/valid.jsonl")
+echo "=== Prism Coder 4B v43 — Local MLX Training ==="
+echo "  Model:       $MODEL"
+echo "  Train rows:  $TRAIN_ROWS"
+echo "  Valid rows:  $VALID_ROWS"
+echo "  Adapter out: $ADAPTER_DIR"
+echo "  Log:         $LOG"
+echo ""
+
+# ── Hyperparameters ──────────────────────────────────────────────────────────
+# r=32, alpha=64 (scale=2.0), 16 LoRA layers, seq_len=2048
+# Effective batch = batch_size(4) × grad_accumulation(4) = 16
+# iters=2000 ≈ 0.35 epochs over ~23k train rows
+# LR 1e-4 cosine — same family as v36 1.7B (5e-6 was after warmup; 1e-4 with cosine suits 4B fresh run)
+# val every 100 steps, save every 200 steps
+# Est. time on M5 Max: 3-5h
+
+mkdir -p "$ADAPTER_DIR"
+
+mlx_lm.lora \
+    --model "$MODEL" \
+    --train \
+    --data "$DATA_DIR" \
+    --adapter-path "$ADAPTER_DIR" \
+    --resume-adapter-file "$ADAPTER_DIR/adapters.safetensors" \
+    --num-layers 16 \
+    --batch-size 2 \
+    --grad-checkpoint \
+    --iters 250 \
+    --val-batches 25 \
+    --learning-rate 3e-5 \
+    --steps-per-report 10 \
+    --steps-per-eval 100 \
+    --save-every 200 \
+    --max-seq-length 2048 \
+    --clear-cache-threshold 0.7 \
+    --seed 2027 \
+    2>&1 | tee -a "$LOG"
+
+echo ""
+echo "✅ Training complete. Adapter at: $ADAPTER_DIR"
+echo "Val loss should be checked in $LOG before merging."
+echo ""
+echo "Next: python3 merge_4b_v43.py  (then re-eval: python3 bfcl_eval.py --model prism-coder:4b-v43)"