lychee-gpt

🍒 Lychee GPT

<!-- Provide a quick summary of what the model is/does. -->

Lychee GPT is a state-of-the-art conversational AI model, fine-tuned on a massive 31B parameter foundation model. Developed by MX LLMS, it delivers exceptional performance in both Bengali and English, featuring a massive 256K context window and stable generation of 11,000+ output tokens in a single pass.

---

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

Lychee GPT is the flagship model of MX LLMS, created to push the boundaries of long-context and long-output text generation for Bengali and English speakers. It is optimized for advanced instruction-following, complex reasoning, and dynamic conversational AI.

• Developed by: MX LLMS & MD Mushfiqur Rahim
• Funded by: Open-source community
• Shared by: Hugging Face Hub
• Model type: Causal Language Model (CLM) with 4-bit QLoRA fine-tuning
• Language(s): Bengali (bn), English (en)
• License: Apache 2.0
• Finetuned from model: A 31B parameter foundation model

Model Sources

• Repository: mahin1234/lychee-gpt
• Demo: Coming soon on Hugging Face Spaces

---

Uses

Direct Use

This model is designed for:

• Chatbots & Virtual Assistants (Bengali & English)
• Long-form content generation (articles, stories, reports)
• Code assistance & reasoning tasks
• Educational and research purposes

Downstream Use

Can be fine-tuned further for specific domains like legal, medical, or customer support.

Out-of-Scope Use

• High-risk decision-making (medical, legal, financial).
• Malicious content generation (hate speech, misinformation, harassment).
• Generating content without human oversight where factual accuracy is critical.

---

Bias, Risks, and Limitations

Lychee GPT inherits biases from its base foundation model. While fine-tuning on curated datasets reduces some harmful outputs, users should:

• Use with human oversight in sensitive applications.
• Avoid relying solely on the model for critical decisions.
• Be aware that the model may produce hallucinations or outdated information.

Recommendations

Users (both direct and downstream) should implement content moderation and validation layers when deploying in production. Always verify factual claims generated by the model.

---

How to Get Started with the Model

Use the code below to load and run Lychee GPT instantly. The LoRA adapter will auto-detect the base model.

**Installation:**https://colab.research.google.com/#fileId=https://huggingface.co/mahin1234/lychee-gpt.ipynb kaggle.com/notebooks/new?accelerator=nvidiaTeslaT4&src=https://huggingface.co/mahin1234/lychee-gpt.ipynb

Training Data

The model was fine-tuned on a custom conversational dataset curated from publicly available sources, including Wikipedia (Bengali & English), open-source GitHub repositories, and curated prompt-response pairs from open foundation models. All data was filtered for quality and cultural relevance.