From 322778f3ddf078fa5e2fd9f8c88e109d020499a3 Mon Sep 17 00:00:00 2001 From: "dingbin.wang" Date: Wed, 21 Oct 2026 17:30:25 +0800 Subject: [PATCH] convert_mmbert --- convert_mmbert.py | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+) create mode 100644 convert_mmbert.py diff --git a/convert_mmbert.py b/convert_mmbert.py new file mode 100644 index 0000000..34bed1e --- /dev/null +++ b/convert_mmbert.py @@ -0,0 +1,22 @@ +#!/usr/bin/env python3 +"""convert_mmbert.py — mmBERT pytorch_model.bin → Laya DecisionModel 格式""" +import torch, os, shutil, sys +from safetensors.torch import save_file + +base = sys.argv[1] if len(sys.argv) > 1 else "/tmp/base" +sd = torch.load(os.path.join(base, "pytorch_model.bin"), map_location="cpu", weights_only=True) +out = {} +dropped = 0 +for k, v in sd.items(): + if k.startswith("cls."): + dropped += 1 + continue + out["encoder." + k] = v.to(torch.float32) +save_file(out, os.path.join(base, "model.safetensors"), metadata={"format": "pt"}) +os.remove(os.path.join(base, "pytorch_model.bin")) +os.makedirs(os.path.join(base, "tokenizer"), exist_ok=True) +for fn in ["tokenizer.json", "tokenizer_config.json", "special_tokens_map.json"]: + src = os.path.join(base, fn) + if os.path.exists(src): + shutil.copy(src, os.path.join(base, "tokenizer", fn)) +print(f"[convert] encoder键 {len(out)} 个, 丢弃mlm头 {dropped} 个")