README.md

metadata

title: RunAsh Chat
emoji: 🪶
colorFrom: pink
colorTo: blue
sdk: docker
pinned: true
license: mit
app_port: 3080
thumbnail: >-
  https://cdn-uploads.huggingface.co/production/uploads/6380f0cd471a4550ff258598/ASSYdCAc5M0h_eiuoZNd3.png
short_description: RunAsh-Chat - All-In-One AI Conversations

🚀 RunAsh-Chat: A LibreChat-Inspired Open-Source Conversational AI

Built for freedom, powered by openness.

“LibreChat, but better — faster, smarter, and fully yours.”

---

Model Description

RunAsh-Chat is an open-source, instruction-tuned large language model designed to replicate and enhance the conversational capabilities of the popular LibreChat ecosystem — while introducing improved reasoning, safety, and multi-turn dialogue handling.

Built upon the Mistral-7B or Llama-3-8B base architecture (depending on variant), RunAsh-Chat is fine-tuned on a curated dataset of high-quality, human-aligned conversations, code assistance prompts, and ethical safety filters. It is optimized for use in self-hosted AI chat interfaces like LibreChat, Ollama, Text Generation WebUI, and local LLM APIs.

Unlike many closed or commercial alternatives, RunAsh-Chat is 100% free to use, modify, and deploy — even commercially — under the Apache 2.0 license.

Key Features

✅ LibreChat-Ready: Seamless drop-in replacement for models used in LibreChat deployments
✅ Multi-Turn Context: Excellent memory of conversation history (up to 8K tokens)
✅ Code & Math Ready: Strong performance on programming, logic, and quantitative reasoning
✅ Safety-Enhanced: Built-in moderation to avoid harmful, biased, or toxic outputs
✅ Lightweight & Fast: Optimized for CPU/GPU inference with GGUF, AWQ, and GPTQ support
✅ Multilingual: Supports English, Spanish, French, German, Portuguese, Russian, Chinese, and more

---

Model Variants

Variant	Base Model	Quantization	Context Length	Link
RunAsh-Chat-v1.0-Mistral-7B	Mistral-7B-v0.1	Q4_K_M GGUF	8K	🤗 Hugging Face
RunAsh-Chat-v1.0-Llama3-8B	Llama-3-8B-Instruct	Q4_K_S GGUF	8K	🤗 Hugging Face
RunAsh-Chat-v1.0-Mistral-7B-AWQ	Mistral-7B-v0.1	AWQ (4-bit)	8K	🤗 Hugging Face

💡 Tip: Use GGUF variants for CPU/Apple Silicon; AWQ/GPTQ for NVIDIA GPUs.

---

Usage Examples

Using with Hugging Face transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "runash-ai/RunAsh-Chat-v1.0-Mistral-7B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are RunAsh-Chat, a helpful assistant."},
    {"role": "user", "content": "Explain quantum computing in simple terms."}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(response)

Using with Ollama

ollama pull runash-chat:7b
ollama run runash-chat "What's the capital of Canada?"

Using with LibreChat

1. Download the GGUF model file (e.g., RunAsh-Chat-v1.0-Mistral-7B.Q4_K_M.gguf)
2. Place it in your models/ folder
3. In config.yml:

   model: "RunAsh-Chat-v1.0-Mistral-7B"
   provider: "ollama"  # or "local"
   

---

Training Data & Fine-Tuning

RunAsh-Chat was fine-tuned using a hybrid dataset including:

• Alpaca and Alpaca-CoT datasets
• OpenAssistant conversations
• Self-instruct and Dolly data
• Human-curated chat logs from open-source AI communities
• Ethical filtering: Removed toxic, biased, or harmful examples using rule-based and model-based moderation

Fine-tuning was performed using LoRA with QLoRA for memory efficiency, on 4× A100 40GB GPUs over 3 epochs.

---

Limitations & Ethical Considerations

⚠️ Not a replacement for human judgment — always validate outputs for critical applications.
⚠️ May hallucinate facts, especially in niche domains — verify with trusted sources.
⚠️ Bias mitigation is ongoing — while trained for fairness, residual biases may persist.
⚠️ Not designed for medical/legal advice — consult professionals.

RunAsh-Chat is not a general-purpose AI agent. It is intended for educational, personal, and non-commercial research use — though commercial use is permitted under Apache 2.0.

---

License

This model is released under the Apache License 2.0 — the same as Mistral and Llama 3. You are free to:

• Use it commercially
• Modify and redistribute
• Build derivative models

Attribution is appreciated but not required.

“LibreChat inspired us. We built something better — and gave it back to the community.”

---

Citation

If you use RunAsh-Chat in your research or project, please cite:

@software{runash_chat_2024,
  author = {RunAsh AI Collective},
  title = {RunAsh-Chat: A LibreChat-Inspired Open-Source Chat Model},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/runash-ai/RunAsh-Chat-v1.0-Mistral-7B}
}

---

Community & Support

🔗 GitHub: https://github.com/runash-ai/runash-chat
💬 Discord: https://discord.gg/runash-ai
🐞 Report Issues: https://github.com/runash-ai/runash-chat/issues
🚀 Contribute: We welcome fine-tuning datasets, translations, and optimizations!

---

Acknowledgments

We gratefully acknowledge the work of:

• Mistral AI for Mistral-7B
• Meta for Llama 3
• The LibreChat community for inspiring accessible AI
• Hugging Face for open model hosting and tools

---

RunAsh-Chat — Because freedom shouldn’t come with a price tag.
Made with ❤️ by the RunAsh AI Collective

---