Presented by:
Trained: 4 x A100 for 17.5 hours on 125k samples
Sponsored by: Garg (@g4rg)
About L3-Aethora-15B v2:
L3 = Llama3
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.
Quants:
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
