diff --git a/train/STATE.md b/train/STATE.md
index 2ed9c24..0d4a9b8 100644
--- a/train/STATE.md
+++ b/train/STATE.md
@@ -80,3 +80,16 @@ Training runs on HF Jobs with Unsloth, not on the Mac. No `mlx_lm` training.
   Stronger check for later: log the Unsloth step-0 loss, or compare the Mac loss after a longer run (the first 200 steps).
 - Full run estimate: 748 steps x 30.5 s = about 6.3 h = about 16 USD (range 12-21 USD, step time varies with document length).
   Not started. Needs Kral's go and the alpha decision.
+
+## Overfit test and full run (2026-10-04, Opus 5.5 decision: alpha 32, cost limit 25 USD)
+
+- Overfit test (job 6ac28c41404719ba3764f7fc): one train document (ZCL_DEMO_ABAP_STRUCTURES, 2503 tokens, train row 18),
+  alpha 32, lr 1e-3, 30 steps, no warmup, 5.8 s/step, about 0.4 USD. Unsloth loss on that document: 0.798 -> 0.000234 (-99.97 %).
+- Mac with the converted adapter (scale 2.0): 0.811 (no adapter) -> 0.012 (-98.5 %). Conversion confirmed (a broken
+  conversion would stay near 0.8). The remaining gap is the base mismatch (bnb nf4 vs MLX affine 4-bit).
+- Full run started: job 6ac28e19fbc85ba6823a0eef, a100-large, 748 steps, rank 16, alpha 32, lr 5e-5 cosine, warmup 30,
+  timeout 8 h (max 20 USD). Valid loss logged at step 0 (log line `EVAL`), checkpoints and valid loss every 200 steps in
+  `erhankeseli/abap-stage1-adapter-full/step<N>/` (with `eval.json`). Expected about 6.3 h.
+- Rule: at step 200 convert the checkpoint (`train/peft_to_mlx.py`), measure the valid loss on the Mac
+  (`runs/stage1/adapter_test_valid_loss.log` command, about 20 min), compare the relative drop with the GPU value.
+  Stop the job (`hf jobs cancel`) if they differ.
diff --git a/train/hf_train.py b/train/hf_train.py
index c4df384..e66fcac 100644
--- a/train/hf_train.py
+++ b/train/hf_train.py
@@ -19,6 +19,10 @@ ap.add_argument("--epochs", type=float, default=2.0)
 ap.add_argument("--rank", type=int, default=16)
 ap.add_argument("--alpha", type=int, default=16)
 ap.add_argument("--lr", type=float, default=5e-5)
+ap.add_argument("--train-file", default="train.jsonl")
+ap.add_argument("--eval-file", default="valid.jsonl")
+ap.add_argument("--warmup-steps", type=int, default=30)
+ap.add_argument("--save-every", type=int, default=0, help="upload the adapter to <out>/step<N> every N steps")
 ap.add_argument("--max-seq-length", type=int, default=16384)
 a = ap.parse_args()
 
@@ -35,26 +39,45 @@ model = FastLanguageModel.get_peft_model(
                     "in_proj_qkv", "in_proj_z", "out_proj"],
     use_gradient_checkpointing="unsloth", random_state=20261003)
 
-ds = load_dataset(a.data, data_files={"train": "train.jsonl", "valid": "valid.jsonl"}, token=tok_hf)
+ds = load_dataset(a.data, data_files={"train": a.train_file, "valid": a.eval_file}, token=tok_hf)
 cfg = SFTConfig(
     output_dir="out", per_device_train_batch_size=1, per_device_eval_batch_size=1,
     gradient_accumulation_steps=1, num_train_epochs=a.epochs, max_steps=a.max_steps,
-    learning_rate=a.lr, lr_scheduler_type="cosine", warmup_steps=min(30, max(1, a.max_steps // 3)) if a.max_steps > 0 else 30,
+    learning_rate=a.lr, lr_scheduler_type="cosine", warmup_steps=a.warmup_steps,
     optim="adamw_8bit", weight_decay=0.0, logging_steps=1, eval_strategy="no", save_strategy="no",
     max_length=a.max_seq_length, dataset_text_field="text", packing=False, seed=20261003, report_to="none")
 tr = SFTTrainer(model=model, processing_class=tok, train_dataset=ds["train"], eval_dataset=ds["valid"], args=cfg)
 
+from transformers import TrainerCallback  # noqa: E402
+from huggingface_hub import HfApi  # noqa: E402
+api = HfApi(token=tok_hf)
+
+
+class SaveCb(TrainerCallback):
+    def on_step_end(self, args, state, control, **kw):
+        n = state.global_step
+        if a.save_every and n % a.save_every == 0:
+            model.save_pretrained(f"ckpt{n}")
+            api.upload_folder(folder_path=f"ckpt{n}", repo_id=a.out, repo_type="model", path_in_repo=f"step{n}")
+            ev = tr.evaluate()
+            print("EVAL", json.dumps({"step": n, "eval_loss": ev.get("eval_loss")}), flush=True)
+            api.upload_file(path_or_fileobj=json.dumps({"step": n, "eval_loss": ev.get("eval_loss")}).encode(),
+                            path_in_repo=f"step{n}/eval.json", repo_id=a.out, repo_type="model")
+
+
+tr.add_callback(SaveCb())
+ev0 = tr.evaluate()
+print("EVAL", json.dumps({"step": 0, "eval_loss": ev0.get("eval_loss")}), flush=True)
 t0 = time.time()
 tr.train()
 train_s = time.time() - t0
 steps = tr.state.global_step
 ev = tr.evaluate()  # valid loss of the adapter, for the conversion check
-info = {"steps": steps, "train_seconds": round(train_s, 1), "sec_per_step": round(train_s / max(steps, 1), 2),
+info = {"eval_loss_step0": ev0.get("eval_loss"), "steps": steps, "train_seconds": round(train_s, 1), "sec_per_step": round(train_s / max(steps, 1), 2),
         "eval_loss": ev.get("eval_loss"), "rank": a.rank, "alpha": a.alpha, "lr": a.lr,
         "peak_gpu_gb": round(__import__("torch").cuda.max_memory_allocated() / 2**30, 1)}
 print("RESULT", json.dumps(info))
 model.save_pretrained("adapter")
 json.dump(info, open("adapter/job_result.json", "w"))
-from huggingface_hub import HfApi  # noqa: E402
-HfApi(token=tok_hf).upload_folder(folder_path="adapter", repo_id=a.out, repo_type="model")
+api.upload_folder(folder_path="adapter", repo_id=a.out, repo_type="model")
 print("PUSHED", a.out)