Upload oss-jev-train for hf jobs uv run

configs/jobs-a10g-scale-10k.yaml

@@ -0,0 +1,24 @@
+# Data scaling sweep: full Hub train (~7909 ≈ ~10k pack). v0.1.1 fix recipe.
+model_id: Qwen/Qwen2.5-0.5B-Instruct
+mode: one_token
+max_length: 1024
+batch_size: 2
+learning_rate: 2.0e-5
+# ~3955 steps/epoch @ bs=2 on 7909 train; 14000 ≈ ~3.5 epochs
+max_steps: 14000
+# null / omit = use all train rows (Hub currently 7909)
+max_train_rows: null
+seed: 42
+fp16: false
+bf16: true
+logging_steps: 100
+save_total_limit: 2
+report_to: none
+use_soft_ce: true
+balance_noul: true
+constrained_codes: true
+noul_n_frac: 0.50
+noul_hard_ce: false
+noul_loss_weight: 1.0
+noul_n_sample_weight: 1.0
+noul_logit_bias_n: 0.0

configs/jobs-a10g-scale-1k.yaml

@@ -0,0 +1,26 @@
+# Data scaling sweep: ~1k train rows. v0.1.1 fix recipe (NOT noul-fix).
+# Constrained CE + soft CE + equal noul Y/N balance. Explicitly disables
+# v0.1.2 defaults (noul_n_frac=0.8, hard CE, loss weights).
+model_id: Qwen/Qwen2.5-0.5B-Instruct
+mode: one_token
+max_length: 1024
+batch_size: 2
+learning_rate: 2.0e-5
+# ~500 steps/epoch @ bs=2 on 1000 train; 1800 ≈ ~3.6 epochs (match fix)
+max_steps: 1800
+max_train_rows: 1000
+seed: 42
+fp16: false
+bf16: true
+logging_steps: 25
+save_total_limit: 2
+report_to: none
+use_soft_ce: true
+balance_noul: true
+constrained_codes: true
+# v0.1.1 equal Y/N among noul (not 0.80 heavy-N)
+noul_n_frac: 0.50
+noul_hard_ce: false
+noul_loss_weight: 1.0
+noul_n_sample_weight: 1.0
+noul_logit_bias_n: 0.0

configs/jobs-a10g-scale-3k.yaml

@@ -0,0 +1,23 @@
+# Data scaling sweep: ~3k train rows. v0.1.1 fix recipe (NOT noul-fix).
+model_id: Qwen/Qwen2.5-0.5B-Instruct
+mode: one_token
+max_length: 1024
+batch_size: 2
+learning_rate: 2.0e-5
+# ~1500 steps/epoch @ bs=2 on 3000 train; 5400 ≈ ~3.6 epochs
+max_steps: 5400
+max_train_rows: 3000
+seed: 42
+fp16: false
+bf16: true
+logging_steps: 50
+save_total_limit: 2
+report_to: none
+use_soft_ce: true
+balance_noul: true
+constrained_codes: true
+noul_n_frac: 0.50
+noul_hard_ce: false
+noul_loss_weight: 1.0
+noul_n_sample_weight: 1.0
+noul_logit_bias_n: 0.0

oss_jev_train/__init__.py

@@ -1,6 +1,6 @@
 """OSS Jev one-token / decision-head training (HF Jobs ready)."""
 
-__version__ = "0.1.2"
+__version__ = "0.1.3"
 SCHEMA_VERSION = "oss-jev-distill-v1"
 # Default smoke base; prefer manifest.base_tokenizer_hint for real runs.
 DEFAULT_MODEL_ID = "Qwen/Qwen2.5-0.5B-Instruct"

oss_jev_train/train.py

@@ -230,9 +230,35 @@ def build_argparser() -> argparse.ArgumentParser:
         default=None,
         help="Serving prior: add this to N logit (default 2.5)",
     )
+    p.add_argument(
+        "--max_train_rows",
+        type=int,
+        default=None,
+        help="Subsample train JSONL to this many rows (seed from cfg)",
+    )
     return p
 
 
+def subsample_examples(
+    examples: list,
+    max_rows: int | None,
+    seed: int,
+) -> list:
+    """Deterministic train subsample for data-size sweeps."""
+    if max_rows is None or int(max_rows) <= 0:
+        return examples
+    n = int(max_rows)
+    if n >= len(examples):
+        return examples
+    import random
+
+    rng = random.Random(int(seed))
+    idx = list(range(len(examples)))
+    rng.shuffle(idx)
+    chosen = sorted(idx[:n])
+    return [examples[i] for i in chosen]
+
+
 def merge_config(args: argparse.Namespace) -> dict[str, Any]:
     cfg: dict[str, Any] = {
         "model_id": DEFAULT_MODEL_ID,
@@ -260,6 +286,7 @@ def merge_config(args: argparse.Namespace) -> dict[str, Any]:
         "noul_n_sample_weight": 2.0,
         "noul_hard_ce": True,
         "noul_logit_bias_n": 2.5,
+        "max_train_rows": None,
     }
     if args.config:
         with open(args.config, encoding="utf-8") as f:
@@ -275,6 +302,7 @@ def merge_config(args: argparse.Namespace) -> dict[str, Any]:
         ("seed", "seed"),
         ("device", "device"),
         ("hub_model_id", "hub_model_id"),
+        ("max_train_rows", "max_train_rows"),
     ):
         val = getattr(args, attr, None)
         if val is not None:
@@ -338,6 +366,12 @@ def run_dry(paths: dict[str, Path], cfg: dict[str, Any]) -> int:
     from transformers import AutoTokenizer
 
     examples = load_jsonl(paths["train"])
+    n_full = len(examples)
+    examples = subsample_examples(
+        examples,
+        cfg.get("max_train_rows"),
+        int(cfg.get("seed", 42)),
+    )
     check_manifest(paths)
     tok = AutoTokenizer.from_pretrained(cfg["model_id"], trust_remote_code=True)
     assert_single_token_codes(examples, tok)
@@ -350,6 +384,8 @@ def run_dry(paths: dict[str, Path], cfg: dict[str, Any]) -> int:
                 "model_id": cfg["model_id"],
                 "mode": cfg["mode"],
                 "n_train": len(examples),
+                "n_train_full": n_full,
+                "max_train_rows": cfg.get("max_train_rows"),
                 "train_path": str(paths["train"]),
                 "output_dir": str(paths["output_dir"]),
                 "use_soft_ce": bool(cfg.get("use_soft_ce", True)),
@@ -385,6 +421,17 @@ def train_one_token(
 
     check_manifest(paths)
     examples = load_jsonl(paths["train"])
+    n_full = len(examples)
+    examples = subsample_examples(
+        examples,
+        cfg.get("max_train_rows"),
+        int(cfg.get("seed", 42)),
+    )
+    print(
+        f"train rows={len(examples)} (full={n_full} "
+        f"max_train_rows={cfg.get('max_train_rows')})",
+        flush=True,
+    )
     eval_examples = (
         load_jsonl(paths["eval"]) if paths["eval"].is_file() else None
     )