README.md

---
language:
  - en
license: cc-by-nc-sa-4.0
library_name: onnxruntime
tags:
  - ner
  - pos
  - dependency-parsing
  - token-classification
  - sequence-classification
  - multi-task
  - onnx
  - int8
  - quantized
  - distilroberta
  - experimental
  - deprecated
datasets:
  - conll2003
  - universal_dependencies
pipeline_tag: token-classification
model-index:
  - name: kniv-distilroberta-nlp-en
    results:
      - task:
          type: token-classification
          name: Named Entity Recognition
        dataset:
          type: conll2003
          name: CoNLL-2003
        metrics:
          - type: f1
            value: 0.909
            name: NER F1
      - task:
          type: token-classification
          name: Part-of-Speech Tagging
        dataset:
          type: universal_dependencies
          name: UD English EWT
        metrics:
          - type: accuracy
            value: 0.976
            name: POS Accuracy
      - task:
          type: token-classification
          name: Dependency Parsing
        dataset:
          type: universal_dependencies
          name: UD English EWT
        metrics:
          - type: accuracy
            value: 0.880
            name: UAS
          - type: accuracy
            value: 0.866
            name: LAS
---

> **⚠️ DEPRECATED — Experimental Model**
>
> This model is an early experimental release from the kniv cascade research
> program and is **no longer maintained**. It predates the current
> 5-head cascade architecture (POS, NER, DEP, SRL, CLS) and the
> bottom-up layer-selective training methodology that produces our
> current production teacher.
>
> **Use the current production model instead:**
> [`dragonscale-ai/kniv-deberta-nlp-base-en-large`](https://huggingface.co/dragonscale-ai/kniv-deberta-nlp-base-en-large)
>
> The current model offers significantly better quality across all tasks,
> includes Semantic Role Labeling and Dialog Act Classification heads,
> and has reproducible benchmarks against standard public test sets.
>
> This repository is preserved for reproducibility and historical reference.
> No further updates, bug fixes, or evaluation runs are planned.

---

# kniv-distilroberta-nlp-en

Multi-task NLP model for English that performs NER, POS tagging, dependency parsing, and sentence classification in a **single forward pass**. Exported as ONNX with INT8 quantization for fast CPU inference.

> **License note:** This model is **CC BY-NC-SA 4.0** (non-commercial) because it was trained on CoNLL-2003 (research use only) and DailyDialog (CC BY-NC-SA). For a commercially licensed version, see the upcoming `kniv-deberta-v3-nlp-en` models trained on the [kniv-corpus-en](https://huggingface.co/datasets/dragonscale-ai/kniv-corpus-en) dataset.

## Model Details

| Field | Value |
|-------|-------|
| **Developer** | [dragonscale-ai](https://huggingface.co/dragonscale-ai) |
| **Base model** | `distilroberta-base` (82M parameters) |
| **Format** | ONNX (FP32 + INT8 quantized) |
| **License** | CC BY-NC-SA 4.0 |
| **Language** | English |
| **Repository** | [rustic-ai/kniv-nlp-models](https://github.com/rustic-ai/kniv-nlp-models) |

### Intended Use

Embedded NLP pipeline for the [uniko](https://github.com/rustic-ai/uniko) cognitive memory system. Designed for fast, multi-task inference on CPU — a single model call extracts entities, POS tags, dependency parse, and sentence type simultaneously.

**Not intended for:** Production commercial use (due to NC license). Use the kniv-deberta-v3 models instead.

## Architecture

Shared DistilRoBERTa encoder with four linear classification heads:

```
Input text
    │
    ▼
┌──────────────────┐
│  DistilRoBERTa   │
│  (6 layers, 768) │
└──┬──┬──┬──┬──────┘
   │  │  │  │
   ▼  ▼  ▼  ▼
  NER POS Dep CLS
```

| Head | Labels | Training Data | Metric |
|------|--------|---------------|--------|
| **NER** | 9 (BIO: PER, ORG, LOC, MISC) | CoNLL-2003 | F1 = 90.9% |
| **POS** | 17 (UPOS tagset) | UD English EWT v2.14 | Acc = 97.6% |
| **Dep** | 1,440 (dep2label composite tags) | UD English EWT v2.14 | UAS = 88.0%, LAS = 86.6% |
| **CLS** | 7 (statement, question, command, ...) | Bootstrap labels | F1 = 78.0% |

### dep2label Encoding

Dependency parsing is reformulated as token classification using [Strzyz et al. (2019)](https://aclanthology.org/N19-1077/). Each token receives a label like `+1@nsubj@VERB` encoding the head direction, dependency relation, and head POS.

## Evaluation Results

### NER (CoNLL-2003 Test)

| Entity | Precision | Recall | F1 |
|--------|-----------|--------|-----|
| PER | 95.6% | 95.9% | **95.7%** |
| LOC | 91.4% | 93.5% | **92.4%** |
| ORG | 88.7% | 89.4% | **89.0%** |
| MISC | 78.4% | 82.8% | **80.5%** |
| **Overall** | **90.1%** | **91.6%** | **90.9%** |

### POS (UD English EWT Test)

**Accuracy: 97.6%** (24,491 / 25,094 tokens correct)

Top confusions: PROPN↔NOUN (229), NOUN↔ADJ (51), ADJ↔VERB (18)

### Dependency Parsing (UD English EWT Test)

| Metric | Score |
|--------|-------|
| **UAS** | **88.0%** |
| **LAS** | **86.6%** |

Performance by sentence length:
- Short (<10 tokens): 88.8%
- Medium (10-20): 89.7%
- Long (20-40): 87.2%
- Very long (40+): 84.1%

### Sentence Classification

| Class | Precision | Recall | F1 |
|-------|-----------|--------|-----|
| statement | 99.9% | 99.9% | 99.9% |
| question | 99.4% | 100% | 99.7% |
| command | 100% | 91.7% | 95.7% |
| filler | 94.3% | 97.1% | 95.7% |
| acknowledgment | 100% | 66.7% | 80.0% |
| greeting | 60.0% | 100% | 75.0% |

## Model Files

| File | Size | Description |
|------|------|-------------|
| `model.onnx` | 316 MB | FP32 ONNX model |
| `model-int8.onnx` | **79 MB** | INT8 quantized (recommended) |
| `label_maps.json` | 31 KB | NER/POS/Dep/CLS label vocabularies |
| `tokenizer.json` | 3.4 MB | DistilRoBERTa tokenizer |

### FP32 vs INT8

| | FP32 | INT8 |
|---|---|---|
| **Size** | 316 MB | **79 MB** (75% smaller) |
| **Latency (CPU)** | 74 ms | **35 ms** (2.1x faster) |
| **NER prediction match** | — | 98.6% |
| **POS prediction match** | — | 98.6% |
| **CLS prediction match** | — | 100% |

## Usage

### Python (ONNX Runtime)

```python
import json
import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer

# Load model and tokenizer
session = ort.InferenceSession("model-int8.onnx")
tokenizer = AutoTokenizer.from_pretrained(".")

with open("label_maps.json") as f:
    labels = json.load(f)

# Run inference
text = "Apple reported $394B revenue in Q4."
inputs = tokenizer(text, return_tensors="np", padding="max_length",
                   max_length=128, truncation=True)

outputs = session.run(None, {
    "input_ids": inputs["input_ids"],
    "attention_mask": inputs["attention_mask"],
})

# outputs[0] = ner_logits, [1] = pos_logits, [2] = dep_logits, [3] = cls_logits

# Extract NER predictions
ner_preds = outputs[0][0].argmax(axis=-1)
tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
for tok, pred in zip(tokens, ner_preds):
    if pred > 0:  # non-O
        print(f"  {tok} → {labels['ner_labels'][pred]}")

# Sentence classification
cls_pred = labels["cls_labels"][outputs[3][0].argmax()]
print(f"  CLS: {cls_pred}")
```

### Rust (ONNX Runtime)

```rust
let session = ort::Session::builder()?.with_model("model-int8.onnx")?;

let encoding = tokenizer.encode("Apple reported $394B revenue.", true);
let outputs = session.run(inputs)?;

let ner_logits = &outputs["ner_logits"];  // [1, seq_len, 9]
let pos_logits = &outputs["pos_logits"];  // [1, seq_len, 17]
let dep_logits = &outputs["dep_logits"];  // [1, seq_len, 1440]
let cls_logits = &outputs["cls_logits"];  // [1, 7]
```

## Training

```bash
git clone https://github.com/rustic-ai/kniv-nlp-models
cd kniv-nlp-models

# This model was trained with the original pipeline (before kniv corpus):
python models/distilroberta-nlp-en/prepare_data.py
python models/distilroberta-nlp-en/train.py
python -m shared.export_onnx \
    --model-dir outputs/distilroberta-nlp-en/best \
    --encoder distilroberta-base \
    --output-dir onnx-output/distilroberta-nlp-en
```

## Limitations

- **Non-commercial license** — trained on CoNLL-2003 (research only) and DailyDialog (CC BY-NC-SA)
- **4 NER types only** — PER, ORG, LOC, MISC (the newer kniv models have 18 types)
- **English only** — no multilingual support
- **128 token max** — sentences longer than 128 subword tokens are truncated
- **CLS labels are bootstrapped** — rule-based heuristics, not gold-standard annotations
- **dep2label decoding depends on POS** — POS errors cascade into dependency errors

## Citation

```bibtex
@misc{kniv-distilroberta-nlp-en,
  title={kniv-distilroberta-nlp-en: Multi-task NLP Model for English},
  author={Dragonscale AI},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/dragonscale-ai/kniv-distilroberta-nlp-en}
}

@inproceedings{strzyz-etal-2019-viable,
    title = "Viable Dependency Parsing as Sequence Labeling",
    author = "Strzyz, Michalina and Vilares, David and G{\'o}mez-Rodr{\'\i}guez, Carlos",
    booktitle = "Proceedings of NAACL",
    year = "2019"
}
```

## See Also

- **Dataset:** [dragonscale-ai/kniv-corpus-en](https://huggingface.co/datasets/dragonscale-ai/kniv-corpus-en) — 657K sentences, CC BY-SA 4.0
- **Commercial models (upcoming):**
  - `kniv-deberta-v3-large-nlp-en` — 304M teacher, 18 NER types, 9 CLS labels
  - `kniv-deberta-v3-nlp-en` — 44M distilled student
- **Source code:** [rustic-ai/kniv-nlp-models](https://github.com/rustic-ai/kniv-nlp-models)