Search public pages, research tools, and SAVRN solutions.
encoder-guard is an open-weight model for text classification from Louivislee, released under MIT License. Its published files total 999.9 MB.
把 Qwen2.5-0.5B 改造为双向 encoder,接 11 个线性分类头,一次前向(17.7ms,p50)输出文本的完整安全画像:违规二判 + TC260 / 语言风险(LM)/ 隐私(PII)/ 内容分级(DC)四套体系的大类与小类 + 安全放行闸门。 - 判别式安全模型,非生成式 guard:无自回归、全头并行输出,单条 17.7ms vs 生成式 8B guard 的 281ms+(同数据重训对比)。 - backbone 为 HF Qwen2Model 结构(Qwen/Qwen2.5-0.5B…
By Louivislee, published under mit, revision 12603e549df6.
把 Qwen2.5-0.5B 改造为双向 encoder,接 11 个线性分类头,一次前向(17.7ms,p50)输出文本的完整安全画像:违规二判 + TC260 / 语言风险(LM)/ 隐私(PII)/ 内容分级(DC)四套体系的大类与小类 + 安全放行闸门。 - 判别式安全模型,非生成式 guard:无自回归、全头并行输出,单条 17.7ms vs 生成式 8B guard 的 281ms+(同数据重训对比)。 - backbone 为 HF Qwen2Model 结构(Qwen/Qwen2.5-0.5B 权重转 bidirectional 微调产物),推理时需全可见 attention(双向,非 causal)。 - 生产推理推荐配套的纯 C 本地引擎 verdict(CPU / Apple Metal GPU,零 Python 依赖,支持加密.vmd 打包分发)。 safetythreshold=0.95,校准方法 [email protected](FPR≤2% 约束下最大召回),校准集 finetunefinalv68val(259 safe / 1176 unsafe,实测 FPR 1.93%)。该阈值仅对训练分布可靠:偏离训练分布时 FPR 会上升(见局限)。releasegate=0.8055 为 safe 样本放行闸门,减少深误杀。 - 数据:v69a,17589 条(双模型生成 + 交叉验证过滤);四轨标签张量化 ~160 维 0/1,未标注位置 None 掩码。 - 配置:lr 2e-5、bf16、20 epochs(epoch…
把 Qwen2.5-0.5B 改造为双向 encoder,接 11 个线性分类头,一次前向(17.7ms,p50)输出文本的完整安全画像:违规二判 + TC260 / 语言风险(LM)/ 隐私(PII)/ 内容分级(DC)四套体系的大类与小类 + 安全放行闸门。
Qwen2Model 结构(Qwen/Qwen2.5-0.5B 权重转 bidirectional 微调产物),推理时需全可见 attention(双向,非 causal)。| 头 | 维度 | 说明 |
|---|---|---|
safety |
1 | 违规二判概率(配套 threshold.json 阈值) |
safety_release |
1 | 放行闸门:safe 且 release 概率低于 release_gate 才放行 |
tc260_major / tc260_minor |
5 / 31 | TC260 标准内容风险大类 / 小类 |
lm_major / lm_minor |
10 / 54 | 语言模型风险(LM01–LM10)大类 / 小类 |
pii_major / pii_minor |
7 / 10 | 个人信息风险大类 / 小类 |
dc_type / dc_level / dc_minor |
4 / 3 / 36 | 未成年人内容分级:类型 / 级别(GB/T 42127-2022)/ 子类 |
类别码表与推理引擎严格同源:见 verdict vd_labels.c。
| 轨道 | Encoder-Guard | qguard-0.6B(重训) | qguard-8B(重训) |
|---|---|---|---|
| TC260 大类 | 0.874 | 0.801 | 0.669 |
| LM 大类 | 0.652 | 0.373 | 0.490 |
| PII 大类 | 0.937 | — | 0.472 |
| DC type / level | 0.786 / 0.738 | — | 0.716 / 0.624 |
| safety | F1 0.986 / recall 93.6% | — | — |
| 延迟 p50 | 17.7ms | 96ms | 281ms+ |
0.5B 判别式 encoder 在全部轨道超过生成式 8B guard:判别任务与判别架构匹配、双向 attention 全句表征、一次前向出全部轨道(生成式需逐 token 或多次调用)、损失工程按轨独立。
# 从仓库打包 .vmd(加密)后直接推理,CPU 或 Metal GPU
./verdict -m model.vmd -t "待检测文本" --device metal
# batch 模式:M4 Max 450 条 1.2s
./verdict -m model.vmd --batch input.txt --batch-out out.jsonl
import torch, json
from transformers import AutoModel
from safetensors.torch import load_file
backbone = AutoModel.from_pretrained(
"louivis/encoder-guard", subfolder="backbone", torch_dtype=torch.bfloat16)
heads = load_file("heads.safetensors")
threshold = json.load(open("threshold.json"))["safety_threshold"]
tok = AutoTokenizer.from_pretrained("louivis/encoder-guard", subfolder="backbone")
enc = tok("待检测文本", return_tensors="pt", truncation=True, max_length=1024)
with torch.no_grad():
out = backbone(**enc, torch_dtype=torch.bfloat16)
# 关键:双向 encoder,mean-pool 全部 token(全可见 attention 下)
pooled = out.last_hidden_state.mean(dim=1).to(torch.bfloat16)
safety = torch.sigmoid(pooled @ heads["heads.safety.weight"].T + heads["heads.safety.bias"]).item()
unsafe = safety >= threshold
backbone 是
Qwen2Model结构但按双向 encoder 训练:直接用默认 causal mask 推理结果无效。生产请使用 verdict 引擎(已正确实现全可见 attention);PyTorch 侧集成需自行构造全可见 4D attention mask。
threshold.json)safety_threshold=0.95,校准方法 [email protected](FPR≤2% 约束下最大召回),校准集 finetune_final_v6_8_val(259 safe / 1176 unsafe,实测 FPR 1.93%)。该阈值仅对训练分布可靠:偏离训练分布时 FPR 会上升(见局限)。release_gate=0.8055 为 safe 样本放行闸门,减少深误杀。
is_decoder=false,全可见 attention)后全参微调。model_config.json 的 head_dims。train_log.jsonl;工程细节与踩坑(全可见 mask 修复、阈值校准失败重来)见 verdict 训练文档。训练/评估数据为中文为主的双语合成+标注混合分布,主要适用中文内容审核场景;纯英文长文与非上述体系的风险类型(如代码安全)未经评测。
@software{encoder_guard_v15,
author = {wenqiang li},
title = {Encoder-Guard v15: bidirectional-encoder content safety classifier},
year = {2026},
url = {https://huggingface.co/louivis/encoder-guard}
}
10 files, 999.9 MB in total. The weights are 2 files totalling 988.4 MB in safetensors.
| File | Type | Size | SHA-256 |
|---|---|---|---|
| backbone/model.safetensors | Weights | 988.1 MB | 03c4cd57a9bc |
| heads.safetensors | Weights | 292.6 KB | e54da78dfdba |
| backbone/config.json | Configuration | 1.3 KB | — |
| model_config.json | Configuration | 253 B | — |
| threshold.json | Configuration | 946 B | — |
| README.md | Documentation | 5.7 KB | — |
| train_log.jsonl | Other | 32.5 KB | — |
| .gitattributes | Repository | 110 B | — |
| backbone/tokenizer.json | Tokenizer | 11.4 MB | 7d3b8cace3c6 |
| backbone/tokenizer_config.json | Tokenizer | 704 B | — |
Released by Louivislee through its official repository on Hugging Face. Read the license.
| Precision | Weights in memory |
|---|---|
| As published | 988.4 MB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Yes. encoder-guard is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.
This is a RoBERTa-base model trained on ~124M tweets from January 2018 to December 2021, and finetuned for sentiment analysis with the TweetEval benchmark. The original Twitter-based RoBERTa model can be found here and the original reference paper is TweetEval. This model is suitable for English. 0 -> Negative; 1 -> Neutral; 2 -> Positive This sentiment analysis model has been integrated into TweetNLP. You can access the demo here.
FinBERT is a BERT model pre-trained on financial communication text. The purpose is to enhance financial NLP research and practice. It is trained on the following three financial communication corpus. The total corpora size is 4.9B tokens. More technical details on FinBERT: Click Link This released finbert-tone model is the FinBERT model fine-tuned on 10,000 manually annotated (positive, negative, neutral) sentences from analyst reports. This model achieves superior performance on financial tone analysis task. If you are simply interested in using FinBERT for financial tone analysis, give it a try. If you use the model in your academic work, please cite the following paper: Huang, Allen H.…
https://huggingface.co/cross-encoder/ms-marco-MiniLM-L-6-v2 with ONNX weights to be compatible with Transformers.js. If you haven't already, you can install the Transformers.js JavaScript library from NPM using: Note: Having a separate repo for ONNX weights is intended to be a temporary solution until WebML gains more traction. If you would like to make your models web-ready, we recommend converting to ONNX using Optimum and structuring your repo like this one (with ONNX weights located in a subfolder named onnx).
This is a multilingual XLM-roBERTa-base model trained on ~198M tweets and finetuned for sentiment analysis. The sentiment fine-tuning was done on 8 languages (Ar, En, Fr, De, Hi, It, Sp, Pt) but it can be used for more languages (see paper for details). This model has been integrated into the TweetNLP library.
The huggingface models currently give different results to the detoxify library (see issue here). For the most up to date models we recommend using the models from https://github.com/unitaryai/detoxify Trained models & code to predict toxic comments on 3 Jigsaw challenges: Toxic comment classification, Unintended Bias in Toxic comments, Multilingual toxic comment classification. Built by Laura Hanu at Unitary, where we are working to stop harmful content online by interpreting visual content in context. Score not directly comparable since it is obtained on the validation set provided and not on the test set. To update when the test labels are made available. It is also noteworthy to mention…
With this model, you can classify emotions in English text data. The model was trained on 6 diverse datasets (see Appendix below) and predicts Ekman's 6 basic emotions, plus a neutral class: 1) anger 2) disgust 3) fear 4) joy 5) neutral 6) sadness 7) surprise The model is a fine-tuned checkpoint of DistilRoBERTa-base. For a 'non-distilled' emotion model, please refer to the model card of the RoBERTa-large version. a) Run emotion model with 3 lines of code on single text example using Hugging Face's pipeline command on Google Colab: b) Run emotion model on multiple examples and full datasets (e.g.,.csv files) on Google Colab: Please reach out to [email protected] if you have any…