Genis is a conversational AI language model trained entirely from scratch — no pretrained weights, no fine-tuning on top of existing models. Every parameter was initialized randomly and trained by BotDevelopmentAI on a single laptop GPU.
Model Details
Property
Value
Architecture
GPT-style Transformer
Parameters
71.2M
Layers
10
Hidden size
768
Attention heads
12
Feedforward dim
3072
Context window
256 tokens
Vocabulary
533 tokens (word-level)
Training data
~30,000 conversation pairs
Epochs
25
Hardware
NVIDIA RTX 4070 Laptop GPU (8.6GB VRAM)
Architecture
Rotary Position Embeddings (RoPE) — better positional understanding than fixed embeddings
RMSNorm — faster and more stable than LayerNorm
Causal self-attention with scaled dot product
Weight tying between embedding and output layers
Top-k + nucleus (top-p) sampling for generation
Training Data
blended_skill_talk — general conversation pairs
SQuAD — 15,000 real question and answer pairs
TriviaQA — 10,000 trivia facts and knowledge pairs
Custom self-knowledge pairs — Genis knows its own identity, version, and builder
Installation
Windows (automatic):
Download all files from this repository
Double-click install.bat — automatically installs Python and all dependencies
Double-click genis_v2_run.bat to start chatting
install.bat handles everything — Python installation, GPU detection, and all packages.
Requirements:
Windows 10 or 11
Internet connection (for install only)
2GB free disk space
NVIDIA GPU recommended but not required (works on CPU too)
Usage
python
1import torch
2from genis_v2_model import GenisV2
3from genis_v2_tokenizer import GenisV2Tokenizer
45# Load model6ckpt = torch.load("genis_v2_model.pt", map_location="cpu", weights_only=False)7tok = GenisV2Tokenizer()8model = GenisV2(vocab_size=ckpt["vocab_size"])9model.load_state_dict(ckpt["model_state"])10model.eval()1112# Chat13defchat(text):14 enc = tok.encode_conversation(text)15 tokens = torch.tensor([enc], dtype=torch.long)16 out = model.generate(tokens, max_new=100, temperature=0.8, top_p=0.92)17 gen = out[0,len(enc):].tolist()18return tok.decode(gen).strip()1920print(chat("hello"))21print(chat("who are you"))22print(chat("tell me a fun fact"))
Example Conversations
YOU: who are you
GENIS: i'm genis, a ai built from scratch by botdevelopmentai.
YOU: hi
GENIS: hi there! i'm genis. how can i help you today?
YOU: tell me a fun fact
GENIS: did you know octopuses have three hearts and blue blood? pretty wild right!
Limitations
533 token vocabulary — cannot output words not in the vocab
Limited world knowledge — trained on 30,000 pairs only
May produce incomplete or incoherent sentences on complex topics
Not suitable for production use — this is a research/educational model
Part of the 51 Ecosystem
Genis v2 is part of a fully custom AI ecosystem built from scratch:
51 Voice — Speech to text
Genis v2 — Conversational AI (this model)
Ludy-1o — Text to image generator
All models designed, built, and trained by BotDevelopmentAI on a single consumer laptop GPU.
What's Next
Genis v3 is in development with:
300M parameters
49,408 token BPE vocabulary (SDXL-style)
Wikipedia training data
Knowledge of celebrities, history, science, geography
License
MIT — free to use, modify, and share.
"Built from scratch, not borrowed." — BotDevelopmentAI
