NILKNARFGonzo/single-floppy-346k-merged

A weird model trained on a Pi. This model is a standalone version; no need to download anything else.

Training Details

Category	Specification
Model Architecture	GPT‑2 (decoder‑only Transformer)
Total Parameters	346,104
Model Size (FP32)	1.32 MB (1,387,072 bytes)
Layers (n_layer)	2
Embedding Dimension (n_embd)	92
Attention Heads (n_head)	2
Context Length (n_positions)	1024
Vocabulary Size	502 (byte‑level BPE, tied embeddings)
Training Framework	Oobabooga TextGen WebUI (Transformers + PEFT)
Fine‑Tuning Method	LoRA (Target all linear layers)
LoRA Rank	8
LoRA Alpha	16
Optimizer	AdamW 8‑bit (adamw_bnb_8bit)
Learning Rate	2e‑4 (linear scheduler, 20 warmup steps)
Batch Size	4
Training Dataset	Capybara ShareGPT (random subset, 1.36 MB, 1,453,408 bytes)
Dataset Format	JSON array with "text" field (flattened ShareGPT)
Training Epochs	5
Loss Target	Stopped at ~6.0 (plateaued due to parameter limit)
Hardware	Raspberry Pi 5, 8GB RAM, ARM Cortex‑A76 (CPU only)
Training Time	~1 hour
Tokenizer	Custom byte‑level BPE (trained on dataset)
Tokenizer Special Tokens	[BOS], [EOS], [PAD], [UNK]
Note	This is a proof‑of‑concept model. It fits on a 3.5‑inch floppy disk and was trained from scratch on a Raspberry Pi.

This thing still produces nonsense. Be warned.