Qybera/SalesA-AI-125M

Model Name with Parameters

SalesA AI 125M — Multimodal Mixture-of-Experts Transformer

• Full Name Example:
  SalesA AI 125M (MoE, Multimodal, N.E.N & SalesA Team)

• Breakdown:

  • SalesA AI: Your model’s brand/architecture.
  • 125M: Number of parameters (from your logs: Total parameters: 125,106,858 ≈ 125M).
  • MoE: Mixture-of-Experts.
  • Multimodal: Handles text, vision, audio, code, action.
  • Author: N.E.N (Nthuku Elijah Nzeli) and SalesA Team.

  Created by N.E.N (Nthuku Elijah Nzeli) and SalesA Team

  • Model architecture: SalesAModel
  • Library: Custom (see included code)
  • Parameters: 125M
  • Modalities: Text, Vision, Audio, Code, Action/Robotics
  • MoE: Yes

  This repository contains the SalesA AI model, a modular, extensible, and efficient multimodal transformer with Mixture-of-Experts (MoE) layers.

Why This Naming Convention?

• Clarity: Users immediately know the model’s size and capabilities.
• Discoverability: Easy to search/filter on Hugging Face and other platforms.
• Professionalism: Follows conventions used by top models (e.g., “Qwen2.5-7B-Minivoc-32k” source, “distilbert-base-uncased” source).

Summary Table

---

SalesA AI 125M — Multimodal Mixture-of-Experts Transformer
Created by N.E.N (Nthuku Elijah Nzeli) and SalesA Team

This repository contains the SalesA AI model, a modular, extensible, and efficient multimodal transformer with Mixture-of-Experts (MoE) layers. It supports text, vision, audio, code, and action/robotics tasks.

• Model architecture: SalesAModel
• Library: Custom (see included code)
• Author: N.E.N (Nthuku Elijah Nzeli) and SalesA Team

Model Description

SalesA AI is a lightweight, CPU-optimized, multimodal model with a Mixture-of-Experts (MoE) architecture. It supports:

• Text generation & classification (including financial sentiment)
• Vision (image-to-text, classification)
• Audio (audio-to-text, classification)
• Code generation
• Action prediction for robotics/locomotion
• Ethical and bias-aware outputs

The model is designed for extensibility, ethical deployment, and real-world applications in finance, sales, stock/market analysis, and robotics.

Intended Uses & Limitations

Intended Uses

• Financial news and sentiment analysis
• Sales and market trend analysis
• General text, vision, and audio tasks
• Code generation and completion
• Robotics: action/command prediction from multimodal input
• Research and educational use

Limitations

• Not suitable for high-stakes financial decisions without human oversight
• May not generalize to all languages or domains
• Biases may exist in training data; see bias analysis plots
• Not for commercial use without review (see license)

Datasets Used

• atrost/financial_phrasebank (financial sentiment)
• allenai/prosocial-dialog (dialogue)
• AI-Lab-Makerere/beans (vision)
• garage-bAInd/Open-Platypus (instruction following)

Training Details

• Architecture: Mixture-of-Experts (MoE), multimodal encoders (text, vision, audio)
• Parameters: ~125M
• Hardware: CPU-optimized, trainable on commodity hardware
• Losses: Cross-entropy for classification/generation, multitask loss
• Optimizer: AdamW
• Batch size: 4 (default)
• Epochs: 10 (default)

Evaluation Results

• Financial sentiment (accuracy): 0.85
• Financial sentiment (F1): 0.83
• General text/vision/audio: See per-task metrics in training logs
• Bias/diagnostic plots: See confusion_matrix.png, class_distribution.png, per_class_metrics.png in the model directory

Ethical Considerations & Bias Analysis

• Model includes bias and diagnostic visualizations for transparency
• Not for use in applications requiring guaranteed fairness or absence of bias
• See Hugging Face Model Card Guide for best practices

Files Included

• model.safetensors: Model weights
• config.json: Model configuration
• tokenizer.json, vocab.json, tokenizer.model: Tokenizer files
• merge.txt: Tokenizer merges (placeholder)
• generation_config.json: Generation parameters
• model.safetensors.index.json: Index for sharded weights (placeholder)
• chat_template.jinja: Chat UI template (placeholder)
• training_history.pkl: Training history
• confusion_matrix.png, class_distribution.png, per_class_metrics.png: Diagnostic plots

How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification
# Or use your custom loading code for SalesA AI

Citation

If you use this model, please cite:

@article{Malo2014GoodDO,
  title={Good debt or bad debt: Detecting semantic orientations in economic texts},
  author={P. Malo and A. Sinha and P. Korhonen and J. Wallenius and P. Takala},
  journal={Journal of the Association for Information Science and Technology},
  year={2014},
  volume={65}
}

License

apache-2.0

Contact

For questions or commercial licensing, contact the SalesA Team.