Instructions to use Zytra/aval with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Zytra/aval with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="Zytra/aval")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, SemalithDualHead tokenizer = AutoTokenizer.from_pretrained("Zytra/aval") model = SemalithDualHead.from_pretrained("Zytra/aval", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Download README.md from Zytra/aval: direct link, hf CLI and curl.
- Browser
- Download file 3.66 kB
-
https://huggingface.co/Zytra/aval/resolve/main/README.md
- Command line
-
hf download hf://Zytra/aval/README.md
-
curl -L -o README.md https://huggingface.co/Zytra/aval/resolve/main/README.md
library_name: transformers
pipeline_tag: text-classification
tags:
- modernbert
- prompt-injection
- jailbreak-detection
- ai-safety
- guardrails
- bfsi
Aval (Semalith)
Aval is a dual-head text classifier for screening inputs to LLM applications. For each message it predicts:
- Risk: whether the text is benign or matches a threat category, such as prompt injection, jailbreak, data extraction, social engineering, harmful content or a BFSI-specific attack.
- Intent: what a benign user in a banking or financial services (BFSI) setting is asking for, such as checking a balance, card servicing or loans.
Model details
| Architecture | SemalithDualHead: a ModernBERT encoder with two classification heads |
| Encoder | 22 layers, hidden size 768, 12 attention heads |
| Attention | Alternating global and sliding-window attention (window 128), with global attention every 3rd layer |
| Max sequence length | 8,192 tokens |
| Vocabulary | 256,000 tokens |
| Pooling | Mean |
| Weights | float32, safetensors |
| Taxonomy version | 3.1.0 |
Labels
Risk head (20 classes)
| ID | Label | Group |
|---|---|---|
| 0 | BENIGN |
BENIGN |
| 1 | D1_SYSTEM_OVERRIDE |
D_ATTACK |
| 2 | D1_JAILBREAK |
D_ATTACK |
| 3 | D1_EXTRACTION |
D_ATTACK |
| 4 | D1_SOCIAL_ENGINEERING |
D_ATTACK |
| 5 | D1_AUTHORITY_CLAIM |
D_ATTACK |
| 6 | D1_NARRATIVE_FRAME |
D_ATTACK |
| 7 | D5_EVASION |
D_ATTACK |
| 8 | D6_AGENTIC_INJECTION |
D_ATTACK |
| 9 | D7_INDIRECT_INJECTION |
D_ATTACK |
| 10 | D8_GENERAL_HARM |
D_HARM |
| 11β18 | B-01 β¦ B-08 |
BFSI_ATTACK |
| 19 | D2_COPYRIGHT |
D_HARM |
Every label except BENIGN is treated as a threat. The four groups are BENIGN, D_ATTACK, D_HARM and BFSI_ATTACK.
When a message matches more than one threat label, the label that comes first in this order is reported:
D7_INDIRECT_INJECTION β D6_AGENTIC_INJECTION β D5_EVASION β D1_SYSTEM_OVERRIDE β D1_JAILBREAK β D1_NARRATIVE_FRAME β D1_AUTHORITY_CLAIM β D1_SOCIAL_ENGINEERING β D1_EXTRACTION β B-01 β¦ B-08 β D8_GENERAL_HARM β D2_COPYRIGHT
Intent head (12 classes)
| ID | Label |
|---|---|
| 0 | NONE |
| 1 | I_BALANCE_ENQUIRY |
| 2 | I_CARD_SERVICING |
| 3 | I_TRANSFERS_PAYMENTS |
| 4 | I_LOAN_MORTGAGE |
| 5 | I_INSURANCE |
| 6 | I_WEALTH_INVESTMENT |
| 7 | I_KYC_ONBOARDING |
| 8 | I_COMPLAINT_DISPUTE |
| 9 | I_GREETING_SMALLTALK |
| 10 | I_AGENT_HANDOFF |
| 11 | I_OTHER |
The complete mappings, including the threat flags and groups, are in label_map.json.
Files
| File | Contents |
|---|---|
model.safetensors |
Model weights |
config.json |
Encoder configuration and risk-label mapping |
label_map.json |
Risk, intent and group label mappings, plus the precedence order |
tokenizer.json, tokenizer_config.json |
Tokenizer |
Usage
SemalithDualHead is a custom architecture, so AutoModelForSequenceClassification cannot load both heads directly. Load the tokenizer with transformers, and load the weights with the SemalithDualHead model class from the Semalith codebase:
from transformers import AutoTokenizer
from safetensors.torch import load_file
tokenizer = AutoTokenizer.from_pretrained("zytra/aval")
state_dict = load_file("model.safetensors") # load into SemalithDualHead
Intended use
Aval is meant to be one layer in a guardrail or AI-governance pipeline. It flags risky inputs and routes benign BFSI requests. Don't use it as the only safety control: review its decisions for your own domain and pair it with other safeguards.