Step 400 check: Mac -13.4 % vs GPU -35.0 %; next GPU run bf16 mixed with stage 2; findings in STATE and roadmap

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

docs/yol-haritasi.md

@@ -1,6 +1,6 @@
 # ABAP Danışman Modeli — Yol Haritası
 
-Durum: v20 · 2026-10-04
+Durum: v21 · 2026-10-05
 Şu anki konum: **Adım 1.3**
 
 ## Adım 0 — Çerçeve ve kararlar
@@ -121,6 +121,13 @@ Adım 1.6 tablosundan seçilir. Aday havuzu: sadece Apache 2.0 veya MIT lisansl
 
 Bitti sayılır: tek baz model seçildi, gerekçesi yazıldı.
 
+## Aşama 1 GPU sonucu ve karar (2026-10-05)
+
+- QLoRA (nf4) aşama 1 koşusu ~418. adımda durduruldu. Valid loss, Mac (MLX 4-bit) / GPU (nf4): adım 200 0,758 (-%10,7) / -%26,0; adım 400 0,735 (-%13,4) / -%35,0; baz Mac 0,849.
+- PEFT→MLX dönüştürücü overfit testiyle kanıtlandı (Mac -%98,5). nf4 adaptörü MLX 4-bit baza iyi taşınmıyor.
+- **Karar:** şimdilik GPU koşusu yok. Aşama 1 tek başına eğitilmez; aşama 1 korpusu, aşama 2 yörüngeleri hazır olunca tek bir bf16 LoRA koşusunda karıştırılır. Sonraki GPU koşusu bf16 baz kullanır; öncesinde bf16 bellek testi gerekir. Araç isimleri EPOD ABAP MCP'deki gibi kalır (generic_v0 kullanılmaz). ADT "save failed" yanıtı sunucuda düzeltilemez: proxy sözdizimi kontrolü ekler.
+- HF test ve checkpoint repoları silindi; aşama 1 veri seti kaldı.
+
 ## Adım 3 — Eğitim verisi üretimi
 
 | Alt adım | İş |

train/STATE.md

@@ -105,3 +105,18 @@ Training runs on HF Jobs with Unsloth, not on the Mac. No `mlx_lm` training.
   on Mac) is not equal. The overfit test (one document) passed, so key names, transpose and scale are right.
 - Open: decide what to do. Options: (a) accept the base mismatch and train further; (b) train on a higher-precision base (bf16 LoRA,
   larger GPU) so the Mac conversion matches; (c) evaluate on the GPU with the MLX-equivalent base.
+
+## Findings and next GPU run (2026-10-05)
+
+| Step | GPU valid loss (nf4 base) | GPU drop | Mac valid loss (MLX 4-bit, converted adapter) | Mac drop |
+|---|---|---|---|---|
+| 0 / base | 0.9228 | | 0.849 | |
+| 200 | 0.6830 | -26.0 % | 0.758 | -10.7 % |
+| 400 | 0.6003 | -35.0 % | 0.735 | -13.4 % |
+
+- Step 400 Mac log: `runs/stage1/full_step400_valid_loss.log`. The Mac drop grows with the steps (-10.7 % to -13.4 %), but stays far below the GPU drop.
+- Converter `train/peft_to_mlx.py` is proven (overfit test: Mac -98.5 %, GPU -99.97 %). The gap is the base mismatch: an nf4 adapter does not transfer well to the MLX 4-bit base.
+- Decisions (Kral, 2026-10-05): no more GPU runs now. Stage 1 is not trained alone. The stage 1 corpus is mixed with the stage 2 trajectories in one bf16 LoRA run later. The next GPU run uses a bf16 base (no nf4). A bf16 memory test is needed before it (27B bf16 weights are about 54 GB; check GPU size, sequence length 16384, gradient checkpointing).
+- Tool names stay as in the EPOD ABAP MCP server (generic_v0 not used). The ADT "save failed" response cannot be fixed on the server side; the proxy adds a syntax check instead.
+- HF cleanup done: test, overfit and full adapter repos deleted (step 200 and 400 checkpoints too). Kept: dataset `erhankeseli/abap-stage1-data`. Local adapter folders and `data_overfit` deleted.
+- `BUDGET_LIMIT_USD` = 161 (ledger 66.97 + 35 usage x 2.7), until 12 October.