L3-Aethora-15B v2 is an advanced language model built upon the Llama 3 architecture. It employs state-of-the-art training techniques and a curated dataset to deliver enhanced performance across a wide range of tasks.

(Thank you all for the interest! the model has surpassed 150k downloads on all formats!)

Quants:

GGUF:

@Mradermacher: L3-Aethora-15B-V2-GGUF
@Bullerwins: L3-Aethora-15B-V2-GGUF

IMatrix-GGUF:

@Mradermacher: L3-Aethora-15B-V2-i1-GGUF

GGUF-F16: (both f16.q6 and f16.q5 are smaller than q8 and perform as well as the pure f16)

@MZeroWw: L3-Aethora-15B-V2-GGUF-f16

EXL2:

@Bullerwins: L3-Aethora-15B-V2-EXL2

Training Process:

Base Model: elinas/Llama-3-15B-Instruct-zeroed
Training Duration: 17.5 hours on 4 x A100 GPUs
Training Method: LoRA (Low-Rank Adaptation)
Epochs: 4
Precision: BF16
Sequence Length: 8192 tokens

Model Capabilities:

The goal of L3-Aethora-15B v2 is to have an expanded proficiency across a wide spectrum of tasks with a focus in creative writing:

Creative Writing and Storytelling:
- Generates engaging narratives, poetry, and creative content
- Adapts writing style to various genres and tones
- Assists in plot development and character creation
General Intelligence:
- Engages in detailed discussions on medical topics and scientific concepts
- Explains complex scientific phenomena
- Assists in literature review and hypothesis generation
Instructional and Educational Content:
- Creates comprehensive tutorials and how-to guides
- Explains complex topics with clarity and appropriate depth
- Generates educational materials for various skill levels
Reasoning and Problem-Solving:
- Analyzes complex scenarios and provides logical solutions
- Engages in step-by-step problem-solving across various domains
- Offers multiple perspectives on challenging issues
Contextual Understanding and Adaptability:
- Maintains coherent, context-aware conversations across extended interactions
- Adapts communication style based on the user's preferences and needs
- Handles nuanced queries with appropriate depth and sensitivity

Dataset Creation Process:

The Aether-Lite-V1.8.1 dataset used for training L3-Aethora-15B v2 underwent a rigorous creation and curation process:

Data Collection: Aggregated from 12 diverse high-quality datasets, including:
- jondurbin/airoboros-3.2
- jtatman/medical-sci-instruct-100k-sharegpt
- Doctor-Shotgun/no-robots-sharegpt
- QuietImpostor/Sao10K-Claude-3-Opus-Instruct-15K-ShareGPT
- TheSkullery/WizardLM_evol_instruct_v2_Filtered_Fuzzy_Dedup_ShareGPT
- TheSkullery/Gryphe-Opus-WritingPrompts-merged
- Alignment-Lab-AI/RPGuild-sharegpt-filtered
- And others, providing a rich mix of instruction, creative writing, and specialized knowledge
Data Preprocessing:
- Language Detection: Utilized a FastText language model to ensure English-language content
- Text Sanitization: Cleaned and normalized text, removing or replacing problematic characters
- Phrase Filtering: Removed specific unwanted phrases and content types
Deduplication:
- Implemented advanced fuzzy deduplication with a 95% similarity threshold
- Utilized text embeddings and cosine similarity calculations for efficient comparison
- Removed 16,250 duplicate entries, ensuring dataset uniqueness
Data Balancing:
- Carefully sampled from each source dataset to maintain diversity
- Implemented data shuffling to ensure random distribution of samples

The final dataset comprises 125,119 high-quality, diverse samples, striking a balance between creativity, practical knowledge, and intellectual depth.

The full dataset used has been released to the public and is avalible for all (see presented section), any ideas or recomendations are always welcome to expand on the dataset further