ghostai1/Egypt_Historical7b

Historical Egyptology 7B (GGUF)

Historical Egyptology 7B — a Mistral-7B-Instruct fine-tune infused with the grandeur, mystery, and wisdom of ancient Egypt.
Perfect for immersive historical roleplay, myth retellings, educational narratives, pharaonic decrees, scribe chronicles, and Nile-soaked storytelling.

---

COMPILING DATA WIP check tomm morn should be fully out

✨ Overview

This model was fine-tuned (LoRA r=64, alpha=64, 3 epochs, 860 cleaned examples) on historical Egyptology texts using Unsloth.
It excels at evocative, period-flavored prose — invoking gods (Ra, Anubis, Isis, Thoth), pharaohs, dynasties, mummification rites, pyramid construction, Nile mythology, and the atmosphere of temples and tombs.

Best uses:

• Deep historical fiction & alternate-history tales
• Educational explainers (mummification, Book of the Dead, Old/Middle/New Kingdom)
• Roleplay as pharaoh, high priest, scribe, tomb robber, explorer, deity…
• Creative local assistants with ancient-Egypt personality

Not optimized for:

• Strict modern factual accuracy (flavor > precision)
• Advanced math, coding, or technical reasoning
• Extremely long context without rope/extensions

Training snapshot (2026-02-05 run):

• Base: Mistral-7B-Instruct (~7.4B params)
• LoRA: 167,772,160 trainable params (2.26%)
• Dataset: 860 examples, max seq len 3072
• Epochs: 3 | Steps: 162 | Final loss: 1.7763
• Runtime: ~11.5 hours (1× GPU)

---

📜 Model Details

• Model name: Historical Egyptology 7B
• Base model: Mistral-7B-Instruct
• Fine-tuning: LoRA (merged via Unsloth)
• Parameters: ~7.4 billion
• Context length: 3072 native (tested up to 8192+ in llama.cpp)
• Language: English + ancient Egyptian stylistic terms
• License: MIT (subject to base model license)

---

🗿 Quantized Files

All files are from the same merged fine-tune checkpoint — only quantization level changes.

File	Quant	Bits	Approx. Size	VRAM est. (4k ctx)	Recommendation
egypt-7b-v1.TQ1_0.gguf	TQ1_0	~1	~1.5 GB	< 2 GB	Ultra-low memory (experimental)
egypt-7b-v1.Q2_K.gguf	Q2_K	~2.5	~2.6 GB	~3 GB	Very low RAM
egypt-7b-v1.Q3_K_S.gguf	Q3_K_S	~3.5	~3.0 GB	~3.5 GB	Low-memory sweet spot
egypt-7b-v1.Q3_K_M.gguf	Q3_K_M	~3.8	~3.3 GB	~4 GB	Balanced low-RAM
egypt-7b-v1.Q4_K_S.gguf	Q4_K_S	~4.5	~3.8 GB	~4.5 GB	Good quality / low VRAM
egypt-7b-v1.Q4_K_M.gguf	Q4_K_M	~4.8	~4.1 GB	~5 GB	Default – best overall balance
egypt-7b-v1.Q5_K_S.gguf	Q5_K_S	~5.5	~4.6 GB	~5.5 GB	Higher quality
egypt-7b-v1.Q5_K_M.gguf	Q5_K_M	~5.7	~4.8 GB	~6 GB	Recommended for best quality
egypt-7b-v1.Q6_K.gguf	Q6_K	~6.6	~5.4 GB	~6.5 GB	Very good detail
egypt-7b-v1.Q8_0.gguf	Q8_0	8	~7.1 GB	~8 GB	Near-lossless reference

Quick picks:

• Most users → Q4_K_M or Q5_K_M (great balance on 6–8 GB cards)
• Tight hardware → Q3_K_M or Q4_K_S
• Maximum fidelity → Q8_0

---

🚀 Usage Examples (llama.cpp)

CLI – basic generation

./llama-cli \
  -m egypt-7b-v1.Q4_K_M.gguf \
  -ngl 35 \                        # adjust GPU layers (0 = CPU only)
  -c 4096 \
  --color \
  -p "You are a high priest of Amun-Ra in Karnak during the reign of Thutmose III. A young acolyte asks you to explain the sacred meaning of the benben stone and its connection to creation. Speak in solemn, evocative language." \
  -n 1024