⚠️ 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This model is an early experimental release from the kniv cascade research program and is no longer maintained.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

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This repository is preserved for reproducibility and historical reference.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 dataset.

Model Details

Field	Value
Developer	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

Intended Use

Embedded NLP pipeline for the 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). 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)

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)

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

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

@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 — 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