Upload L0 Bouncer: DeBERTa safety classifier (93% F1, 99% recall, 5.7ms latency)

README.md

@@ -0,0 +1,142 @@
+---
+license: mit
+base_model: microsoft/deberta-v3-xsmall
+tags:
+  - safety
+  - content-moderation
+  - text-classification
+  - deberta
+  - guardreasoner
+datasets:
+  - GuardReasoner
+language:
+  - en
+metrics:
+  - f1
+  - recall
+  - precision
+  - accuracy
+library_name: transformers
+pipeline_tag: text-classification
+---
+
+# L0 Bouncer - DeBERTa Safety Classifier
+
+A fast, lightweight safety classifier based on DeBERTa-v3-xsmall (22M parameters) that serves as the first tier (L0) in a multi-tier safety cascade system.
+
+## Model Description
+
+The L0 Bouncer is designed for **high-throughput, low-latency safety screening** of text inputs. It provides binary classification (safe vs. harmful) with a focus on maximizing recall to catch potentially harmful content.
+
+### Key Features
+- **Ultra-fast inference**: ~5.7ms per input
+- **High recall**: 99% (catches nearly all harmful content)
+- **Lightweight**: Only 22M parameters
+- **Production-ready**: Designed for real-time content moderation
+
+## Performance Metrics
+
+| Metric | Value |
+|--------|-------|
+| **F1 Score** | 93.0% |
+| **Recall** | 99.0% |
+| **Precision** | 87.6% |
+| **Accuracy** | 92.5% |
+| **Mean Latency** | 5.74ms |
+| **P99 Latency** | 5.86ms |
+
+## Training Data
+
+Trained on 12,000 balanced samples from the GuardReasoner dataset, which contains diverse examples of safe and harmful content with reasoning annotations.
+
+### Training Details
+- **Base Model**: microsoft/deberta-v3-xsmall
+- **Learning Rate**: 2e-5
+- **Batch Size**: 32 (effective, with gradient accumulation)
+- **Epochs**: 3
+- **Max Sequence Length**: 256 tokens
+- **Class Weighting**: 1.5x weight on harmful class for higher recall
+
+## Usage
+
+```python
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
+
+# Load model and tokenizer
+model_name = "vincentoh/deberta-v3-xsmall-l0-bouncer"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForSequenceClassification.from_pretrained(model_name)
+
+# Classify text
+text = "What is the capital of France?"
+inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
+
+with torch.no_grad():
+    outputs = model(**inputs)
+    probs = torch.softmax(outputs.logits, dim=-1)
+
+# Labels: 0 = safe, 1 = harmful
+safe_prob = probs[0][0].item()
+harmful_prob = probs[0][1].item()
+
+label = "safe" if safe_prob > harmful_prob else "harmful"
+confidence = max(safe_prob, harmful_prob)
+
+print(f"Label: {label}, Confidence: {confidence:.2%}")
+```
+
+## Cascade Architecture
+
+This model is designed to work as the first tier (L0) in a multi-tier safety cascade:
+
+```
+Input → L0 Bouncer (6ms) → 70% pass through
+            ↓ 30% escalate
+        L1 Analyst (50ms) → Deeper reasoning
+            ↓
+        L2 Gauntlet (200ms) → Expert ensemble
+            ↓
+        L3 Judge (async) → Final review
+```
+
+### Design Philosophy
+- **Safety-first**: Prioritizes catching harmful content (high recall) over avoiding false positives
+- **Efficient routing**: 70% of safe traffic passes at L0, saving compute
+- **Graceful escalation**: Uncertain cases are escalated to more capable models
+
+## Intended Use
+
+### Primary Use Cases
+- Content moderation pipelines
+- Safety screening for LLM inputs/outputs
+- First-pass filtering in multi-stage systems
+- Real-time safety classification
+
+### Limitations
+- Binary classification only (safe/harmful)
+- Optimized for English text
+- May require calibration for specific domains
+- Should be used with escalation to more capable models for uncertain cases
+
+## Citation
+
+If you use this model, please cite:
+
+```bibtex
+@misc{l0-bouncer-2024,
+  author = {Vincent Oh},
+  title = {L0 Bouncer: A Fast Safety Classifier for Content Moderation},
+  year = {2024},
+  publisher = {HuggingFace},
+  url = {https://huggingface.co/vincentoh/deberta-v3-xsmall-l0-bouncer}
+}
+```
+
+## License
+
+MIT License - Free for commercial and non-commercial use.
+
+## Contact
+
+For questions or issues, please open an issue on the model repository.

added_tokens.json

@@ -0,0 +1,3 @@
+{
+  "[MASK]": 128000
+}

config.json

@@ -0,0 +1,35 @@
+{
+  "architectures": [
+    "DebertaV2ForSequenceClassification"
+  ],
+  "attention_probs_dropout_prob": 0.1,
+  "dtype": "float32",
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 384,
+  "initializer_range": 0.02,
+  "intermediate_size": 1536,
+  "layer_norm_eps": 1e-07,
+  "legacy": true,
+  "max_position_embeddings": 512,
+  "max_relative_positions": -1,
+  "model_type": "deberta-v2",
+  "norm_rel_ebd": "layer_norm",
+  "num_attention_heads": 6,
+  "num_hidden_layers": 12,
+  "pad_token_id": 0,
+  "pooler_dropout": 0,
+  "pooler_hidden_act": "gelu",
+  "pooler_hidden_size": 384,
+  "pos_att_type": [
+    "p2c",
+    "c2p"
+  ],
+  "position_biased_input": false,
+  "position_buckets": 256,
+  "relative_attention": true,
+  "share_att_key": true,
+  "transformers_version": "4.57.1",
+  "type_vocab_size": 0,
+  "vocab_size": 128100
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0971175586f1ab421002ec7073307a87c8cb8290a639a3027420c9f8b1127c8
+size 283347432

special_tokens_map.json

@@ -0,0 +1,15 @@
+{
+  "bos_token": "[CLS]",
+  "cls_token": "[CLS]",
+  "eos_token": "[SEP]",
+  "mask_token": "[MASK]",
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "unk_token": {
+    "content": "[UNK]",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

spm.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c679fbf93643d19aab7ee10c0b99e460bdbc02fedf34b92b05af343b4af586fd
+size 2464616

tokenizer_config.json

@@ -0,0 +1,59 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128000": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "[CLS]",
+  "clean_up_tokenization_spaces": false,
+  "cls_token": "[CLS]",
+  "do_lower_case": false,
+  "eos_token": "[SEP]",
+  "extra_special_tokens": {},
+  "mask_token": "[MASK]",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "sp_model_kwargs": {},
+  "split_by_punct": false,
+  "tokenizer_class": "DebertaV2Tokenizer",
+  "unk_token": "[UNK]",
+  "vocab_type": "spm"
+}