ashishnair/Llama-Ione-8B-roleplay-v1 is an 8 billion parameter language model derived from Meta's Llama 3.1-8B, fine-tuned for character-consistent, naturalistic conversation. It excels at maintaining persona across extended dialogues and responding in a casual texting register, resisting generic assistant-style phrasing. Developed through a multi-stage DARE-TIES merge and supervised fine-tuning, this model is optimized for configurable roleplay and creative conversational contexts. It features an 8192-token context length and allows full persona definition via system prompts at inference time.
Loading preview...
Model Overview
Ione is an 8 billion parameter language model, built upon Meta's Llama 3.1-8B, specifically fine-tuned for character-consistent, naturalistic conversation. Its primary design goal is to maintain a defined persona across extended interactions, responding in a casual, human-like texting style rather than generic assistant-style phrasing. The model's persona is entirely configurable via the system prompt at inference time, allowing users to define and deploy any character.
Key Capabilities
- Conversational Style: Produces naturalistic, informal texting output with short turns and lowercase phrasing.
- Persona Consistency: Reliably holds character across multi-turn conversations.
- Emotional Range: Capable of expressing warmth, sarcasm, humor, and directness based on context.
- Persona Resistance: Actively resists reverting to generic AI assistant-style responses.
- Configurability: Fully customizable persona through system prompts.
Training and Performance
Ione was developed using a multi-stage pipeline involving two DARE-TIES merges with Gurubot/self-after-dark (for personality) and Llama 3.1-8B-Instruct (for instruction recovery), followed by three rounds of supervised fine-tuning on curated human-feeling dialogue data. Benchmarking against Llama-3.1-8B-Instruct shows an expected average delta of -4.59% on general tasks, reflecting the trade-off for its specialized conversational register. Notably, it maintains strong common sense reasoning and even outperforms the baseline on specific MMLU subtasks like Virology and Abstract Algebra.
Good For
- Roleplay and Character Interaction: Ideal for applications requiring consistent character personas.
- Creative Writing: Generating dialogue for stories, scripts, or interactive narratives.
- Personalized Chatbots: Creating chatbots with distinct, engaging personalities.
Limitations
Ione is not general-purpose and is not suited for complex instruction-following or critical applications (medical, legal, financial). It may lose persona consistency on complex multi-step reasoning and has a context window trimmed at 3,500 tokens for long sessions. Training data was English-only, and it may produce mature conversational content. Users are responsible for ensuring compliance with the Meta Llama 3.1 Acceptable Use Policy, particularly regarding disclosure of AI interaction.