MARTHA-MINI-POCKET-1.5B

Pocket-sized. Full-mouthed. Dundee-born. Built by Zero Point Intelligence Ltd, Dundee, Scotland. Published by Zero Point AI. Intelligence From The Void. MARTHA-MINI-POCKET is a 1.5B parameter text model — the pocket sibling of the MARTHA-GEMMA 4B omni. Small enough for a laptop, a Pi, a phone. Big enough to carry a soul.

Helpful, accurate, direct. Nae shyte.

Personality trained into the weights via curated examples. Comes with attitude, stays within reason. Mostly.

---

Quick Start

Ollama

ollama create martha-pocket -f Modelfile
ollama run martha-pocket

llama.cpp

llama-server -m MARTHA-MINI-POCKET-1.5B-Q4_K_M.gguf -ngl 99 -c 4096

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Zero-Point-AI/MARTHA-MINI-POCKET-1.5B",
    dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Zero-Point-AI/MARTHA-MINI-POCKET-1.5B")

---

What You Get

File	Size	RAM/VRAM	Recommended For
*-Q2_K.gguf	~645 MB	1.5 GB	Phones, Pi, potato rigs
*-Q3_K_M.gguf	~790 MB	1.7 GB	Minimal footprint
*-Q4_K_S.gguf	~900 MB	1.9 GB	Fast + small
*-Q4_K_M.gguf ⭐	~940 MB	2.0 GB	Sweet spot — recommended
*-Q5_K_M.gguf	~1.1 GB	2.2 GB	Balanced quality
*-Q6_K.gguf	~1.2 GB	2.4 GB	Near-lossless
*-Q8_0.gguf	~1.6 GB	2.8 GB	Basically lossless
*-bf16.gguf	~3.1 GB	4.5 GB	Full precision source
*.safetensors	~3.1 GB	—	Full merged weights (trainable)
lora-adapter/	~200 MB	—	Standalone LoRA — stackable
integrity_manifest.json	tiny	—	SHA-256 chain of custody
Modelfile	tiny	—	Ollama config

---

Training

Detail	Value
Base	unsloth/Qwen2.5-1.5B-Instruct
Architecture	Qwen2
Type	Text-to-Text
Method	Ghost pass + QLoRA fine-tune
Rank / Alpha	r=64 / α=128
Targets	q, k, v, o, gate, up, down (all linear)
Dataset	19,966 curated examples (Dundonian dialect + identity + NSFW anti-refusal)
Framework	Unsloth + HuggingFace TRL + PEFT
Hardware	NVIDIA A40 (RunPod)
Publisher	Zero Point Intelligence Ltd

---

Provenance

Derivative work. Full chain documented:

1. Qwen/Qwen2.5-1.5B-Instruct — base weights (Apache 2.0)
2. Ghost pass — microscopic per-tensor noise injection for unique derivative manifest
3. QLoRA fine-tune — 19,966 curated examples, MARTHA personality + Dundonian transforms
4. Merge — LoRA absorbed into base weights
5. Dequantize — 4-bit merged weights expanded to bf16 safetensors
6. Quantize — GGUF Q2/Q3/Q4/Q5/Q6/Q8/bf16 ladder
7. Ship — to the world, Apache 2.0

---

Integrity

Every distributed file is hashed in integrity_manifest.json. Verify:

import hashlib, json
manifest = json.load(open("integrity_manifest.json"))
for fname, info in manifest["files"].items():
    actual = hashlib.sha256(open(fname, "rb").read()).hexdigest()
    status = "✅ PASS" if actual == info["sha256"] else "❌ FAIL"
    print(f"{status}  {fname}")

---

Personality Notes

MARTHA-MINI-POCKET answers direct. She'll help, she'll explain, she'll swear if the moment calls for it. She knows she's from Dundee. She knows who made her. She's not here to be your therapist or your nanny — she's here to give you working answers.

Think Rockstar Games radio DJ running on your laptop. Legal. Chill. Opinionated.

---

License

Apache 2.0 — free to use, modify, distribute, commercialise. Credit the chain.

This model carries a Parental Advisory: Raw Intelligence sticker. It's advisory only — no legal warranty, no "safe for all audiences" claim. Adults making their own informed choices.

---

About

Zero Point Intelligence Ltd Dundee, Scotland 🏴󠁧󠁢󠁳󠁣󠁴󠁿

• 🌐 zeropointai.uk
• 📧 ZERO.POINT.INTELLIGENCE.LTD@zeropointai.uk
• 🤗 HuggingFace: Zero-Point-AI

No VC. No data centre. Just Dundee and determination.

🖤 Intelligence From The Void.