Aya-In-Brooklyn/spaCy-roberta-workout-entity-model

Low-Latency Voice-to-NLP Recommender (NER) — RoBERTa Fine-Tuned

Fine-tuned roberta-base spaCy model for extracting fitness-related entities from voice-transcribed or free-text user queries. Trained on 30,000+ synthetic ASR-style queries using TransformerListener and TransitionBasedParser architecture.

Model Overview

• —Entities:
• —WORKOUT_TYPE, DURATION, INSTRUCTOR, INTENSITY, GOAL, MOOD
• —Base Model: roberta-base
• —spaCy Components:
• —spacy-transformers.TransformerListener.v1 with RoBERTa upstream
• —spacy.TransitionBasedParser.v2
• —Training Data:
• —30,000+ unique ASR-simulated fitness queries
• —Covers realistic variations in duration, mood, intensity, goal, instructor, and type
• —Evaluation:
• —Overall NER F1 Score: 99.97%
• —Token Accuracy: 100%
• —Evaluation performed on a held-out dev set

Architecture

Input Processing:

1. 1.spaCy Tokenization (custom whitespace and rule-based split)
2. 2.RoBERTa Tokenizer (subword-level BPE tokenization)
3. 3.Embedding Layer (lookup table to 768-dimensional dense vectors)
4. 4.Transformer Encoder (multi-head self-attention and feedforward blocks)
5. 5.TransformerListener:
6. 6.Pools subword embeddings to align with spaCy tokens
7. 7.reduce_mean.v1 pooling used for subword combinations
8. 8.TransitionBasedParser (NER Decoder):
9. 9.Projects token vectors (768 → 128 hidden units)
10. 10.Applies maxout layers and span transition modeling
11. 11.Predicts BIO-tag labels per token

Training Configuration:

• —Optimizer: Adam with weight decay
• —Learning Rate: 1e-4
• —Weight Decay (L2 Regularization): 0.01
• —Dropout: 0.15
• —Maxout Pieces: 2
• —Mixed Precision: Enabled (AMP)

Usage

Install the required packages:

pip install spacy spacy-transformers

Load the fine-tuned model:

import spacy

nlp = spacy.load("Aya-In-Brooklyn/fitness_entity_extractor_ner_roberta_finetuned")

doc = nlp("Find me a 45 minute yoga class with Alex.")

for ent in doc.ents:
    print(ent.text, ent.label_)

Example Output:

45 minute -> DURATION
yoga -> WORKOUT_TYPE
Alex -> INSTRUCTOR

Training Details

• —Loss Function: Categorical Cross-Entropy over entity labels
• —Span Construction:
• —BIO (Begin, Inside, Outside) tagging
• —State transitions between spans handled with a transition-based system
• —Evaluation Metrics:
• —Precision (P), Recall (R), F1 Score (F) for each entity type
• —Entity-level (not token-level) evaluation
• —Batcher:
• —batch_by_words.v1, size tolerance 20%
• —Maximum batch size: 2048 words

Training was tracked and logged with MLflow for full reproducibility.

Model Evaluation (Detailed)

• —Overall NER F1: 99.97%
• —Inference Speed: ~5950 tokens/sec on single GPU

Source Structure

• —config_filled.cfg: Full model training configuration
• —voice_assistant/models/entity/model-last/: Saved fine-tuned spaCy model
• —voice_assistant/data/entity_data/: Training and development datasets
• —MLflow artifacts: Logged metrics and parameters for training reproducibility

Deployment

Model is ready to be pushed to Hugging Face Hub or integrated into production systems via:

import spacy
nlp = spacy.load("Aya-In-Brooklyn/fitness_entity_extractor_ner_roberta_finetuned")

Notes

• —This model assumes ASR-style, noisy voice queries.
• —Designed for realistic, imperfect, user-facing voice systems.
• —No manual rule-based extraction is applied; pure end-to-end learning from labeled spans.