From 1976230250fd89dd569e3df461fb99ccf4391dbc Mon Sep 17 00:00:00 2001
From: iomgaa <iomgaaycz@gmail.com>
Date: Sat, 18 Jul 2026 10:14:31 -0400
Subject: [PATCH] =?UTF-8?q?=E5=B1=821:=20=E6=8D=9F=E5=A4=B1=E6=8E=A2?=
 =?UTF-8?q?=E9=92=88=E8=84=9A=E6=9C=AC=E2=80=94=E2=80=94=E9=A2=84=E8=AE=AD?=
 =?UTF-8?q?=E7=BB=83=E6=A8=A1=E5=9E=8B=E8=B5=B0=E7=AE=A1=E7=BA=BF=E9=80=90?=
 =?UTF-8?q?=E8=A1=8C=E7=AE=97=20CE=EF=BC=8C=E5=8C=BA=E5=88=86=E6=95=B0?=
 =?UTF-8?q?=E6=8D=AE=E9=97=AE=E9=A2=98=E4=B8=8E=E8=AE=AD=E7=BB=83=E7=8E=AF?=
 =?UTF-8?q?=E8=8A=82=E9=97=AE=E9=A2=98?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 scripts/diag_loss_probe.py | 62 ++++++++++++++++++++++++++++++++++++++
 1 file changed, 62 insertions(+)
 create mode 100644 scripts/diag_loss_probe.py

diff --git a/scripts/diag_loss_probe.py b/scripts/diag_loss_probe.py
new file mode 100644
index 0000000..1094c25
--- /dev/null
+++ b/scripts/diag_loss_probe.py
@@ -0,0 +1,62 @@
+"""损失探针：用未训练的预训练模型走完整管线，逐行算 loss（层 1 疑点排查第二步）。
+
+判读（训练日志初始 loss ≈ 7.5）：
+- 探针也 ≈ 7：管线一致，loss 高是数据/模型现实 → 去查数据（垃圾长文、乱码占比）；
+- 探针 ≈ 2-4：管线（本脚本与训练共用）没问题但训练环节另有妖 → 查训练循环差异。
+同时打印 HF 模型内建 CE（同一数学的独立实现）交叉验证 sft_loss。
+
+远程运行（CPU 即可，约 1-2 分钟）：
+    python -u scripts/diag_loss_probe.py
+"""
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from ars_opd.configs import SFTConfig
+from ars_opd.data import SFTCollator, load_sft_dataset
+from ars_opd.trainer import sft_loss
+
+MODEL = "Qwen/Qwen3-0.6B"
+
+cfg = SFTConfig(
+    dataset_path="data/dapo-math-17k-unique.parquet",
+    output_dir="/tmp/diag",
+    subset_size=8,
+    seed=42,
+    teacher_completions_path="data/teacher_completions_dapo1k_minimax-m3.jsonl",
+)
+ds = load_sft_dataset(cfg)
+tok = AutoTokenizer.from_pretrained(MODEL)
+collator = SFTCollator(
+    tok,
+    max_length=cfg.max_length,
+    max_prompt_length=cfg.max_prompt_length,
+    enable_thinking=False,
+)
+model = AutoModelForCausalLM.from_pretrained(MODEL, dtype=torch.float32)
+model.eval()
+
+print(f"{'行':>3} {'sft_loss':>9} {'HF内建CE':>9} {'监督tok':>7}  解答开头")
+total, total_n = 0.0, 0
+for i in range(len(ds)):
+    batch = collator([ds[i]])
+    with torch.no_grad():
+        out = model(
+            input_ids=batch["input_ids"], attention_mask=batch["attention_mask"]
+        )
+        ours, n = sft_loss(
+            out.logits, batch["input_ids"], batch["labels"], batch["attention_mask"]
+        )
+        # 交叉验证：HF 内建损失（labels 传入模型，内部自动移位）与 sft_loss
+        # 是同一数学的两个独立实现，单行 batch 下应当几乎相等
+        hf = model(
+            input_ids=batch["input_ids"],
+            attention_mask=batch["attention_mask"],
+            labels=batch["labels"],
+        ).loss
+    head = ds[i]["messages"][-1]["content"][:40].replace("\n", " ")
+    print(f"{i:>3} {ours.item():>9.3f} {hf.item():>9.3f} {n:>7}  {head}", flush=True)
+    total += ours.item() * n
+    total_n += n
+
+print(f"\n按 token 加权平均: {total / total_n:.3f}（对照训练日志初始 loss ≈ 7.5）", flush=True)