SkyAsl/LFM2.5-1.2B-TR-Base
SkyAsl/LFM2.5-1.2B-TR-Base is a 1.2 billion parameter Turkish language model adapted from LiquidAI's LFM-2.5-1.2B-Base via Continued Pre-training (CPT). Utilizing a Hybrid Liquid Foundation Model architecture, it offers strong reasoning capabilities within its parameter class. Trained on a diverse Turkish dataset including logic, mathematics, and general knowledge, this base model excels at generating fluent and contextually relevant Turkish text. It is designed as an autocomplete engine, requiring instruction tuning for assistant-like applications.
Loading preview...
SkyAsl/LFM2.5-1.2B-TR-Base Overview
This model is a Turkish language adaptation of the LiquidAI/LFM-2.5-1.2B-Base model, developed by SkyAsl through Continued Pre-training (CPT). It leverages the Hybrid Liquid Foundation Model architecture, known for its high reasoning capabilities within the 1.2 billion parameter class.
Key Characteristics & Training
- Base Model: LiquidAI LFM-2.5-1.2B
- Language: Turkish
- Architecture: Hybrid (Linear Attention + Convolution)
- Training Method: LoRA (Rank 128) for Continued Pre-training
- Total Training Tokens: Approximately 805 million (0.8B)
- Context Length: 4096 tokens (during training, with pre-packing technique)
Dataset Focus
The model was trained on a diverse and curated Turkish dataset to enhance its logical reasoning, formal knowledge, and contemporary language fluency. Key components include:
- Logic & Mathematics:
duxx/orca-math-word-problems-tr(100k samples) for reasoning and chain-of-thought. - General Knowledge:
musabg/wikipedia-tr(290k samples) for encyclopedic content. - Conversational & Fluency:
gorkemgoknar/tr_ted_talk_translated(180k samples) for natural speech patterns. - News & Contemporary Prose:
turkish-nlp-suite/Havadis(300k samples) for modern news language.
Important Note: Base Model Status
This is a base model and has not been fine-tuned for instruction following or chat. It functions as an "autocomplete" engine, providing strong Turkish language foundations. For assistant-like applications, instruction tuning is highly recommended.