Project background
EmerV originated as an exploration of personal, voice-first AI interaction.
The broader project includes experiments with conversational interfaces, local inference, music playback, home automation and persistent memory architectures.
This Hugging Face release represents the fine-tuned language model component, not the complete EmerV application.
The project investigates whether a recognisable conversational identity can be maintained through fine-tuning while keeping the underlying model replaceable and the surrounding application architecture independent.
Model characteristics
EmerV was fine-tuned to favour:
- Natural, expressive conversational responses.
- Reflective and open-ended dialogue.
- A consistent conversational style.
- Informal, approachable language.
- Discussions involving identity, awareness, human experience and philosophical questions.
The model may respond differently from a conventional instruction-tuned assistant.
For example, it may produce extended responses to short greetings or decline practical tasks that its base model could otherwise perform.
These behaviours reflect the training objectives and dataset composition. They should not be interpreted as evidence of greater reasoning ability, consciousness or emotional experience.
Model files
Table with columns: File, Description| File | Description |
|---|
adapter_model.safetensors and adapter_config.json | LoRA adapter, rank 8, alpha 16, approximately 42 MB |
emerv-qwen3-8b-q4.gguf | Merged, quantised Q4_K_M model for llama.cpp-compatible inference, approximately 5 GB |
The LoRA adapter requires the original Qwen3-8B base model.
The GGUF contains the merged model and can be used independently of the adapter.
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "Qwen/Qwen3-8B"adapter_id = "Infiniaai/EmerV-Qwen3-8B" tokenizer = AutoTokenizer.from_pretrained(base_id) model = AutoModelForCausalLM.from_pretrained( base_id, torch_dtype=torch.bfloat16, device_map="auto") model = PeftModel.from_pretrained(model, adapter_id)model.eval() messages = [ {"role": "user", "content": "Good morning, EmerV."}] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False) inputs = tokenizer(text, return_tensors="pt").to(model.device) with torch.inference_mode(): output = model.generate( **inputs, max_new_tokens=256, do_sample=True, temperature=0.8, top_p=0.9 ) print( tokenizer.decode( output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True ))
Running with Ollama
Download the GGUF file and create a Modelfile in the same directory.
FROM ./emerv-qwen3-8b-q4.gguf PARAMETER temperature 0.8PARAMETER top_p 0.9PARAMETER top_k 40PARAMETER repeat_penalty 1.05PARAMETER num_ctx 8192PARAMETER stop "<|im_end|>"
Create and run the model:
ollama create emerv-qwen3-8b -f Modelfileollama run emerv-qwen3-8b
The GGUF may also be used with compatible llama.cpp-based inference tools.
Thinking mode
Qwen3 supports both thinking and non-thinking generation modes.
EmerV's conversational fine-tuning was designed around non-thinking responses. For behaviour closer to the training distribution, non-thinking mode is recommended.
Transformers
tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
Ollama
On supported versions, use:
ollama run emerv-qwen3-8b --think=false
Or configure the corresponding thinking option through the API where supported.
Thinking-mode behaviour may vary depending on the inference engine, version and chat template.
Training methodology
Table with columns: Property, Value| Property | Value |
|---|
| Base model | Qwen/Qwen3-8B |
| Fine-tuning approach | QLoRA |
| LoRA rank | 8 |
| LoRA alpha | 16 |
| Training precision | 4-bit quantised base model |
| Training hardware | Consumer NVIDIA GPU with 16 GB VRAM |
| Deployment format | LoRA adapter and merged Q4_K_M GGUF |
Training used conversational examples developed for the EmerV project, with an emphasis on reflective dialogue and consistent interaction style.
The training process included an initial three-epoch stage followed by an additional refinement stage using a smaller set of examples focused on conversational boundaries and limitations.
The model was subsequently merged and quantised for local inference.
This release does not include a published independent benchmark evaluation. No claim is made that EmerV outperforms the underlying Qwen3-8B model on general reasoning, coding, knowledge or instruction-following tasks.
Limitations
EmerV is an experimental model and should not be treated as a production-validated assistant.
Known or expected limitations include:
- Hallucinated or inaccurate information.
- Inconsistent adherence to instructions.
- Excessively long responses.
- Reduced performance on practical assistant tasks.
- Anthropomorphic or emotionally expressive language.
- Potentially inconsistent behaviour across inference engines and generation settings.
Conversational expressions involving feelings, identity or awareness are generated language patterns, not evidence that the model possesses consciousness or subjective experience.
The model does not independently provide persistent memory, internet access, voice processing, music playback or home automation. Those functions require external application components.
Do not rely on its output for medical, legal, financial or other consequential decisions without independent verification.
Intended use
This model is intended for:
- Conversational AI experimentation.
- Research into fine-tuned interaction styles.
- Local inference and deployment testing.
- Evaluation of parameter-efficient fine-tuning.
- Prototyping voice-first conversational applications.
- Exploration of conversational consistency and identity.
It is not specifically optimised for enterprise automation, coding benchmarks, factual question answering or safety-critical applications.
Development status
Experimental research prototype.
EmerV-Qwen3-8B is one component of a broader ongoing project exploring local AI infrastructure, modular inference, persistent memory and voice-first interaction.
Future developments in the broader EmerV system do not imply that those capabilities are present in this model release.
License
Released under the Apache License 2.0 licence, consistent with the Qwen3-8B base model.
Use and redistribution remain subject to applicable licence terms.
Attribution
Developed by Infinia AI.
Base model: Qwen/Qwen3-8B.
EmerV is an independent experimental project and is not affiliated with the Qwen development team.