Basically-Human-4B
Description
Basically-Human-4B is a 4B parameter roleplaying language model built on Qwen3 4B, fine-tuned for immersive, emotionally resonant character interaction. Designed to emulate the flow and nuance of real human dialogue, it excels at staying in persona, crafting believable responses, and driving dynamic storytelling. With support for ChatML formatting and alignment focused on realism and consistency, Basically-Human-4B is a compact but capable choice for text-based roleplay, NPC simulation, and interactive fiction.
Instruct Format
Basically-Human-4B uses the ChatML instruction format and was trained under the Qwen3 chat template. Messages are structured with system, user, and assistant roles, each enclosed in <|im_start|> and <|im_end|> tokens, ensuring compatibility with applications that rely on structured conversational data. This format enables clear role delineation and seamless multi-turn interactions, making the model highly effective for in-character dialogue and roleplay scenarios.
Procedure
The pretrained base model was fine-tuned on instruct and roleplaying data in a single step. Once the supervised fine-tuning was complete, a Kahneman-Tversky optimization was applied as a final alignment step.
Data
The model was trained on a diverse collection of instruction and roleplaying data, including the following sets:
- PocketDoc/Dans-MemoryCore-CoreCurriculum-Small
- NewEden/Kalo-Opus-Instruct-22k-Refusal-Murdered
- Epiculous/Synthstruct-Gens-v1.1-Filtered-n-Cleaned
- NewEden/Gryphe-Sonnet-3.5-35k-Subset
- Nitral-AI/GU_Instruct-ShareGPT
- Nitral-AI/Medical_Instruct-ShareGPT
- AquaV/Resistance-Sharegpt
- AquaV/US-Army-Survival-Sharegpt
- Gryphe/Sonnet3.5-SlimOrcaDedupCleaned
- Epiculous/SynthRP-Gens-v1.1-Filtered-n-Cleaned
- ResplendentAI/bluemoon
- hardlyworking/openerotica-freedomrp-sharegpt-system
- MinervaAI/Aesir-Preview
- anthracite-core/c2_logs_32k_v1.1
- Nitral-AI/Creative_Writing-ShareGPT
- PJMixers/lodrick-the-lafted_OpusStories-Story2Prompt-ShareGPT
- NewEden/Opus-accepted-hermes-rejected-shuffled
Quantizations
We provide GGUF quantized versions of Basically-Human-4B:
- Downloads last month
- 5