AlfredJames/jobbert-zh-v6a

JobBERT-zh v6a (human reference set)

This repository is the human-reference continuation of JobBERT-zh (artifact Gold150). It is an additional version, not a replacement of the V4 hybrid 2601 encoder row.

Checkpoint	Hub	Test	Typed exact F1
V4 (paper-main encoder)	AlfredJames/jobbert-zh	V4 hybrid 2601 + jieba	0.4331
v6a B2 (this repo)	AlfredJames/jobbert-zh-v6a	Human reference set	0.5536±0.0054 (n=3 sample SD)

Do not rank 0.5536 against 0.4331, official Qwen SOP extract 0.1724, or Gold v2 ChatGPT 0.6365 in one table. SD is a three-seed sample SD, not a test-set confidence interval. The human reference freeze is not in Zenodo v0.1.1.

• Code: https://github.com/AlfredJamesLi/chinese-skillspan-benchmark
• V4 weights (unchanged): https://huggingface.co/AlfredJames/jobbert-zh
• Human-reference notes: https://github.com/AlfredJamesLi/chinese-skillspan-benchmark/blob/main/data/gold150/README.md

---

What this repository contains

Same 3M DAPT encoder as AlfredJames/jobbert-zh. Only the CRF head is continued from the released V4 crf/best.pt on Silver-plus v6a B2. The published teacher files are v6a_nocross (2,150 train / 169 frozen dev). A 2,156-row train list is an earlier v6a count before the nocross dedupe; do not treat 2,156 as the released file.

Path	Role
config.json, model.safetensors, tokenizer.json, tokenizer_config.json	Encoder (identical SHA to AlfredJames/jobbert-zh)
crf/best.pt	Default load path = seed 42 (human-reference exact 0.5544; closest to the 3-seed mean)
crf/seed42/best.pt	Seed 42 — human-reference exact 0.5544 / relaxed 0.6917
crf/seed43/best.pt	Seed 43 — human-reference exact 0.5479 / relaxed 0.6794
crf/seed44/best.pt	Seed 44 — human-reference exact 0.5587 / relaxed 0.6959

Paper cell 0.5536±0.0054 / relaxed 0.6890±0.0085 is the mean ± sample SD of those three seeds. Loading only seed 42 is not the published mean.

File	Bytes	SHA-256
encoder model.safetensors	406,730,376	ed2130f680d0aa9691d081a516963da252449746ef7e78818050e3039b6ccf3b
crf/seed42/best.pt	409,169,393	7bea1f284b666570fd542c1fc7c9f1101840a11be16b4c04cb303047a8df7b7d
crf/seed43/best.pt	409,169,393	275b97da7a94d2d42c3ae0fe8e88c4fe2dec857831e0dc958578cb593abc06d8
crf/seed44/best.pt	409,169,393	3d6d708b8633264aed47911eaeaa317dc052dff1bea34c814165ae42e68cc626

Init CRF (not in this repo): V4 AlfredJames/jobbert-zh crf/best.pt, SHA-256 d98814cb954036f885e1247c2c50c66adbc5a280750f0228bfc3df077f2fc98c.

---

Training

• Encoder frozen as the released 3M JobBERT-zh (ckpt65000).
• CRF continued with AdamW (2\times10^{-5}), weight decay 0.01, batch 16, at most 6 epochs, patience 2, 10% warmup.
• Checkpoint = argmax frozen-dev typed exact.
• Labels: Silver-plus teacher B2 (not B1). Isolation: sentence-level (extended).
• Scorer: cnss-lskt-1.2.0. Human-reference freeze SHA-256 ca8db0bc386c24543fec845d42ea8e24883eb67b8ce75129ac1768c5c0310fd0 (data/gold150_test.jsonl).
• No jieba snap on this human-reference cell (character CRF vs human overlay).

There is no AutoModelForTokenClassification export.

---

Loading

import torch
from torch import nn
from torchcrf import CRF
from huggingface_hub import hf_hub_download
from transformers import AutoModel, AutoTokenizer

REPO = "AlfredJames/jobbert-zh-v6a"

class BertCRF(nn.Module):
    def __init__(self, model_dir: str, n_labels: int = 9, dropout: float = 0.1):
        super().__init__()
        self.encoder = AutoModel.from_pretrained(model_dir)
        self.dropout = nn.Dropout(dropout)
        self.emissions = nn.Linear(self.encoder.config.hidden_size, n_labels)
        self.crf = CRF(n_labels, batch_first=True)

tok = AutoTokenizer.from_pretrained(REPO)
model = BertCRF(REPO)
# Default = seed 42. For the paper mean, evaluate seed42/43/44 and average.
crf_path = hf_hub_download(REPO, "crf/best.pt")
model.load_state_dict(torch.load(crf_path, map_location="cpu"))

The paper class is BertCRF in scripts/train_cn_roberta_crf.py.

---

Out of scope

• Replacing AlfredJames/jobbert-zh (0.4331 on V4 hybrid 2601)
• Applicant screening or ESCO concept-ID prediction
• Treating 0.5536 as a test CI, or as an independent re-test of the same systems on 2,601 IDs (human-reference IDs sit inside hybrid 2601)
• Peel-rerun heads (0.5509); those did not beat 0.5536 and are not this dump

---

Licence

license: other until job-advertisement text rights are confirmed. GitHub LICENSE Apache-2.0 is for repository software, not these weights. Same authors and funding (NSSFC 21BGL142) as JobBERT-zh.