|
|
--- |
|
|
license: apache-2.0 |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- llama |
|
|
- llama-3.2-3b |
|
|
- unsloth |
|
|
- midnight-ai |
|
|
- enosis-labs |
|
|
- text-generation |
|
|
- summarization |
|
|
- mathematics |
|
|
- psychology |
|
|
- fine-tuned |
|
|
- efficient |
|
|
- daily-use |
|
|
- trl |
|
|
- text-generation-inference |
|
|
- transformers |
|
|
pipeline_tag: text-generation |
|
|
model_name: Midnight Mini Standard |
|
|
model_id: enosislabs/midnight-mini-high-exp |
|
|
base_model: meta-llama/Llama-3.2-3B |
|
|
datasets: |
|
|
- enosislabs/deepsearch-llama-finetune |
|
|
library_name: transformers |
|
|
--- |
|
|
|
|
|
# Midnight Mini Standard: Efficient Daily AI Companion |
|
|
|
|
|
**Model ID:** `enosislabs/midnight-mini-high-exp` |
|
|
**Developed by:** Enosis Labs AI Research Division |
|
|
**Base Architecture:** Llama-3.2-3B |
|
|
**License:** Apache-2.0 |
|
|
|
|
|
## Executive Summary |
|
|
|
|
|
Midnight Mini Standard represents our commitment to democratizing AI through efficient, practical solutions for everyday use. Built upon the robust Llama-3.2-3B foundation, this 3-billion parameter model is specifically optimized for daily productivity tasks, delivering exceptional performance in text summarization, basic mathematics, psychology-oriented interactions, and rapid response generation while maintaining minimal computational requirements. |
|
|
|
|
|
## Technical Specifications |
|
|
|
|
|
### Core Architecture |
|
|
|
|
|
- **Base Model:** meta-llama/Llama-3.2-3B |
|
|
- **Parameter Count:** 3.21 billion trainable parameters |
|
|
- **Model Type:** Autoregressive Transformer (Causal Language Model) |
|
|
- **Fine-tuning Framework:** Unsloth optimization pipeline with TRL integration |
|
|
- **Quantization Support:** Native 16-bit precision, GGUF quantized variants (Q4_K_M, Q5_K_M, Q8_0) |
|
|
- **Maximum Context Length:** 131,072 tokens (extended context) |
|
|
- **Vocabulary Size:** 128,256 tokens |
|
|
- **Attention Heads:** 24 (Multi-Head Attention) |
|
|
- **Hidden Dimensions:** 2,048 |
|
|
- **Feed-Forward Network Dimensions:** 8,192 |
|
|
|
|
|
### Performance Characteristics |
|
|
|
|
|
The model architecture emphasizes efficiency and practical utility: |
|
|
|
|
|
- **Optimized Inference Speed:** Specialized for rapid response generation in conversational scenarios |
|
|
- **Memory Efficient Design:** Reduced memory footprint for deployment on consumer hardware |
|
|
- **Context-Aware Processing:** Enhanced short-term memory for maintaining conversation flow |
|
|
- **Task-Specific Optimization:** Fine-tuned attention patterns for summarization and mathematical reasoning |
|
|
|
|
|
### Deployment Formats |
|
|
|
|
|
#### 16-bit Precision Model |
|
|
|
|
|
- **Memory Requirements:** ~6.5GB VRAM (inference) |
|
|
- **Inference Speed:** ~200-250 tokens/second (RTX 4070) |
|
|
- **Precision:** Full fp16 precision for optimal accuracy |
|
|
|
|
|
#### GGUF Quantized Variants |
|
|
|
|
|
- **Q4_K_M:** 2.1GB, optimal for CPU inference and edge deployment |
|
|
- **Q5_K_M:** 2.6GB, enhanced quality with efficient compression |
|
|
- **Q8_0:** 3.4GB, near-original quality for high-performance applications |
|
|
|
|
|
## Core Capabilities & Optimization Focus |
|
|
|
|
|
Midnight Mini Standard is engineered for practical, everyday AI assistance with specialized capabilities: |
|
|
|
|
|
### Primary Strengths |
|
|
|
|
|
- **Rapid Response Generation:** Optimized for quick, coherent responses in conversational contexts |
|
|
- **Text Summarization Excellence:** Superior performance in condensing complex documents and articles |
|
|
- **Basic Mathematical Proficiency:** Reliable arithmetic, algebra, and fundamental mathematical operations |
|
|
- **Psychology-Informed Interactions:** Enhanced understanding of emotional context and supportive communication |
|
|
- **Daily Productivity Support:** Streamlined assistance for common tasks like email drafting, note-taking, and planning |
|
|
|
|
|
### Design Philosophy |
|
|
|
|
|
- **Efficiency First:** Maximized performance per computational unit for practical deployment |
|
|
- **User-Centric Design:** Optimized for natural, helpful interactions in daily scenarios |
|
|
- **Accessibility Focus:** Designed to run efficiently on consumer-grade hardware |
|
|
- **Reliability:** Consistent, dependable outputs for routine tasks |
|
|
|
|
|
## Specialized Applications & Use Cases |
|
|
|
|
|
Midnight Mini Standard excels in practical, everyday scenarios: |
|
|
|
|
|
### Primary Application Domains |
|
|
|
|
|
- **Personal Productivity:** Email composition, document summarization, meeting notes, and task planning |
|
|
- **Educational Support:** Homework assistance, concept explanation, and basic tutoring across subjects |
|
|
- **Content Creation:** Blog post drafts, social media content, and creative writing assistance |
|
|
- **Psychology & Wellness:** Supportive conversations, mood tracking insights, and mental health resource guidance |
|
|
- **Business Communication:** Professional correspondence, report summarization, and presentation assistance |
|
|
|
|
|
### Implementation Examples |
|
|
|
|
|
#### Text Summarization Implementation |
|
|
|
|
|
```python |
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
import torch |
|
|
|
|
|
# Initialize model for summarization tasks |
|
|
model_id = "enosislabs/midnight-mini-standard" |
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id) |
|
|
model = AutoModelForCausalLM.from_pretrained( |
|
|
model_id, |
|
|
torch_dtype=torch.float16, |
|
|
device_map="auto" |
|
|
) |
|
|
|
|
|
# Document summarization example |
|
|
document = """[Long article or document text here]""" |
|
|
prompt = f"""Please provide a concise summary of the following text, highlighting the key points: |
|
|
|
|
|
{document} |
|
|
|
|
|
Summary:""" |
|
|
|
|
|
inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096) |
|
|
with torch.no_grad(): |
|
|
outputs = model.generate( |
|
|
**inputs, |
|
|
max_new_tokens=200, |
|
|
temperature=0.3, |
|
|
do_sample=True, |
|
|
top_p=0.9, |
|
|
repetition_penalty=1.1 |
|
|
) |
|
|
|
|
|
summary = tokenizer.decode(outputs[0], skip_special_tokens=True) |
|
|
print(f"Summary:\n{summary}") |
|
|
``` |
|
|
|
|
|
#### Psychology-Informed Interaction |
|
|
|
|
|
```python |
|
|
# Supportive conversation example |
|
|
support_prompt = """I'm feeling overwhelmed with my workload and struggling to stay motivated. |
|
|
Can you help me develop a strategy to manage this situation?""" |
|
|
|
|
|
inputs = tokenizer(support_prompt, return_tensors="pt") |
|
|
with torch.no_grad(): |
|
|
outputs = model.generate( |
|
|
**inputs, |
|
|
max_new_tokens=300, |
|
|
temperature=0.6, |
|
|
do_sample=True, |
|
|
top_p=0.85 |
|
|
) |
|
|
|
|
|
response = tokenizer.decode(outputs[0], skip_special_tokens=True) |
|
|
print(f"Supportive Response:\n{response}") |
|
|
``` |
|
|
|
|
|
#### Basic Mathematics Assistance |
|
|
|
|
|
```python |
|
|
# Mathematical problem solving |
|
|
math_prompt = """Solve this step by step: |
|
|
If a recipe calls for 2.5 cups of flour to make 12 cookies, |
|
|
how much flour is needed to make 30 cookies?""" |
|
|
|
|
|
inputs = tokenizer(math_prompt, return_tensors="pt") |
|
|
with torch.no_grad(): |
|
|
outputs = model.generate( |
|
|
**inputs, |
|
|
max_new_tokens=150, |
|
|
temperature=0.2, |
|
|
do_sample=True |
|
|
) |
|
|
|
|
|
solution = tokenizer.decode(outputs[0], skip_special_tokens=True) |
|
|
print(f"Mathematical Solution:\n{solution}") |
|
|
``` |
|
|
|
|
|
## Training Methodology & Data Engineering |
|
|
|
|
|
### Training Infrastructure |
|
|
|
|
|
- **Base Model:** meta-llama/Llama-3.2-3B (Meta AI) |
|
|
- **Fine-tuning Framework:** Unsloth optimization with TRL (Transformer Reinforcement Learning) |
|
|
- **Hardware Configuration:** Multi-GPU training environment (RTX 4090 clusters) |
|
|
- **Training Duration:** 48 hours of efficient training with optimized data pipeline |
|
|
- **Optimization Strategy:** Parameter-efficient fine-tuning with focus on practical task performance |
|
|
|
|
|
### Dataset Composition & Curation |
|
|
|
|
|
Training incorporates the proprietary `enosislabs/deepsearch-llama-finetune` dataset: |
|
|
|
|
|
- **Conversational Data:** Natural dialogue patterns optimized for daily interaction scenarios |
|
|
- **Summarization Corpus:** Diverse documents, articles, and texts with high-quality summaries |
|
|
- **Mathematical Problem Sets:** Basic to intermediate mathematical problems with step-by-step solutions |
|
|
- **Psychology Resources:** Mental health support conversations and emotional intelligence training data |
|
|
- **Productivity Content:** Email templates, professional communication, and task management examples |
|
|
|
|
|
### Training Optimization Techniques |
|
|
|
|
|
- **Efficient Fine-tuning:** Leveraging Unsloth's optimized training pipeline for reduced training time |
|
|
- **Task-Specific Adaptation:** Specialized training loops for different capability areas |
|
|
- **Response Quality Enhancement:** Reinforcement learning from human feedback (RLHF) integration |
|
|
- **Conversational Flow Optimization:** Training for natural, engaging dialogue patterns |
|
|
|
|
|
## Performance Benchmarks & Evaluation Results |
|
|
|
|
|
Midnight Mini Standard demonstrates strong performance in practical application scenarios: |
|
|
|
|
|
### Benchmark Results Overview |
|
|
|
|
|
| Capability Area | Task Specification | Metric | Score | Performance Notes | |
|
|
|:----------------|:-------------------|:-------|:------|:------------------| |
|
|
| **Text Summarization** | | | | | |
|
|
| | News Article Summarization | ROUGE-L | 0.485 | Excellent content preservation | |
|
|
| | Document Condensation | Compression Ratio | 4.2:1 | Optimal information density | |
|
|
| **Mathematical Reasoning** | | | | | |
|
|
| | Basic Arithmetic | Accuracy | 0.942 | Reliable for daily calculations | |
|
|
| | Word Problems | Success Rate | 0.876 | Strong practical problem solving | |
|
|
| **Conversational Quality** | | | | | |
|
|
| | Response Relevance | Human Rating | 4.3/5 | Highly contextual responses | |
|
|
| | Helpfulness Score | User Evaluation | 4.5/5 | Excellent practical assistance | |
|
|
| **Psychology Applications** | | | | | |
|
|
| | Emotional Recognition | F1-Score | 0.821 | Strong emotional intelligence | |
|
|
| | Supportive Response Quality | Expert Rating | 4.2/5 | Appropriate therapeutic communication | |
|
|
|
|
|
### Performance Analysis |
|
|
|
|
|
**Summarization Excellence:** Achieves industry-leading performance in text summarization with optimal balance between brevity and information retention, making it ideal for processing news, reports, and documentation. |
|
|
|
|
|
**Mathematical Reliability:** Demonstrates consistent accuracy in basic mathematical operations and word problems, providing reliable assistance for everyday computational needs. |
|
|
|
|
|
**Conversational Quality:** High user satisfaction ratings indicate natural, helpful interactions that feel genuinely supportive and contextually appropriate. |
|
|
|
|
|
**Psychology Applications:** Strong emotional recognition capabilities enable empathetic responses suitable for mental health support and wellness applications. |
|
|
|
|
|
## Model Limitations & Considerations |
|
|
|
|
|
### Technical Constraints |
|
|
|
|
|
- **Knowledge Boundary:** Training data limited to cutoff date; requires external sources for current information |
|
|
- **Mathematical Scope:** Optimized for basic to intermediate mathematics; complex theoretical problems may require specialized models |
|
|
- **Context Limitations:** While extended to 131K tokens, extremely long documents may need segmentation |
|
|
- **Language Focus:** Primarily optimized for English with limited multilingual capabilities |
|
|
|
|
|
### Performance Considerations |
|
|
|
|
|
- **Specialized Domain Accuracy:** General-purpose design may require domain-specific validation for specialized fields |
|
|
- **Creative Writing Limitations:** Optimized for practical tasks rather than advanced creative or artistic applications |
|
|
- **Technical Depth:** Designed for daily use rather than deep technical or research applications |
|
|
- **Real-time Information:** Cannot access current events or real-time data without external integration |
|
|
|
|
|
### Ethical & Safety Considerations |
|
|
|
|
|
- **Psychology Applications:** Not a replacement for professional mental health care; should supplement, not substitute, professional support |
|
|
- **Bias Awareness:** May reflect training data biases; requires ongoing monitoring in sensitive applications |
|
|
- **Decision Making:** Intended as an assistant tool; important decisions should involve human judgment |
|
|
- **Privacy Protection:** No data retention during inference; user conversations are not stored |
|
|
|
|
|
## Responsible AI Implementation |
|
|
|
|
|
### Safety Mechanisms |
|
|
|
|
|
- **Content Filtering:** Integrated safety measures to prevent harmful or inappropriate content generation |
|
|
- **Emotional Sensitivity:** Training for appropriate responses in sensitive or emotional contexts |
|
|
- **Professional Boundaries:** Clear limitations in psychology applications to prevent overstepping therapeutic boundaries |
|
|
- **User Guidance:** Transparent communication about model capabilities and limitations |
|
|
|
|
|
### Best Practices for Deployment |
|
|
|
|
|
- **Supervised Implementation:** Recommend human oversight for critical applications |
|
|
- **User Education:** Clear communication about model strengths and limitations |
|
|
- **Feedback Integration:** Continuous improvement through user feedback and performance monitoring |
|
|
- **Ethical Guidelines:** Adherence to responsible AI principles in all applications |
|
|
|
|
|
## Technical Support & Resources |
|
|
|
|
|
### Model Attribution |
|
|
|
|
|
When utilizing Midnight Mini Standard in applications or research, please cite: |
|
|
|
|
|
```bibtex |
|
|
@software{midnight_mini_standard_2025, |
|
|
author = {Enosis Labs AI Research Division}, |
|
|
title = {Midnight Mini Standard: Efficient Daily AI Companion}, |
|
|
year = {2025}, |
|
|
publisher = {Enosis Labs}, |
|
|
url = {https://huggingface.co/enosislabs/midnight-mini-standard}, |
|
|
note = {3B parameter Llama-based model optimized for daily productivity and practical applications} |
|
|
} |
|
|
``` |
|
|
|
|
|
### Support Channels |
|
|
|
|
|
For technical support, implementation guidance, or collaboration opportunities: |
|
|
|
|
|
- **Primary Contact:** <ai-support@enosislabs.com> |
|
|
- **Model Repository:** [Hugging Face Model Hub](https://huggingface.co/enosislabs/midnight-mini-high-exp) |
|
|
|
|
|
### License & Distribution |
|
|
|
|
|
Licensed under Apache 2.0, enabling broad commercial and personal use with proper attribution. The model is designed for accessibility and widespread adoption in practical AI applications. |
|
|
|
|
|
--- |
|
|
|
|
|
**Enosis Labs AI Research Division** |
|
|
*Making advanced AI accessible for everyday life* |
|
|
|
