Model Overview
aimeri/spoomples-qwen3-14b-v0.2 is a 14 billion parameter language model based on the Qwen3-14B architecture. It was developed by aimeri through a multi-stage training process involving Continued Pre-Training (CPT) and Supervised Fine-Tuning (SFT). The CPT stage injected domain knowledge from character cards and literary prose, while the SFT stage focused on instruction-following and conversation using a custom chat format called DanChat.
Key Capabilities & Training Focus
- Persona Consistency: Designed to maintain a consistent voice and behavior, exemplified by the 'Olivia' persona (a 31-year-old Brazilian zoologist turned ML hobbyist). This demonstrates the ability to train in specific personas rather than relying solely on prompting.
- Custom Chat Format: Utilizes a unique DanChat format with
<|system|>, <|user|>, <|assistant|>, and <|endoftext|> tokens for structured conversations. - Diverse Training Data: The SFT training data blend includes significant weighting for:
- Roleplay & Creative Writing (28%)
- NSFW content (22%)
- Tasks & Instructions (17%)
- Reasoning & Logic (16%)
- Persona Voice reinforcement (12%)
- Specialized Knowledge (5%)
Intended Use Cases
- Roleplay and Character-Driven Conversation: Its primary strength lies in engaging in character-consistent dialogues.
- Creative and Narrative Writing: Suitable for generating imaginative and story-based content.
- Reasoning and Problem-Solving Tasks: Capable of handling logical and analytical challenges.
- Instruction Following and Tool Use: Can follow instructions and assist with tool-related tasks, though it may not match models specifically optimized for these functions.
Limitations
This is an SFT checkpoint and has not yet undergone DPO (preference alignment), meaning outputs may not always align with user expectations for tone or safety. The model was trained with a specific data mix and custom chat format, so performance with other chat templates may vary. No formal benchmarks have been conducted.