North-ML1/wind-arc-1-5-preview

Wind Arc 1.5 — Preview

An early experimental release from the Wind Arc line. This model is significantly behind Wind Edge 1.6 — treat it as a research artifact, not a usable assistant.

North ML · Wind Edge 1.6 – Geode

⚠️ Preview release. Underperforms Wind Edge 1.6 by roughly 50% across benchmarks. Not recommended for any practical use.

---

Overview

Wind Arc 1.5 Preview exists to document an early training run and architectural experiment. It is not a capable model by any standard measure. Published for transparency and research continuity — the stable Wind Arc line will be a full rebuild.

---

Benchmark Results

Task	Score	Notes
MMLU (boosted)	0.08 / 1.00	Boosted eval with few-shot scaffolding

On MMLU: 0.08 on a boosted protocol means near-random performance even with prompt engineering assistance. Raw 0-shot will score lower. Wind Edge 1.6 scores roughly 2× this across comparable tasks. This model is not competitive.

---

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("arthu1/wind-arc-1-5-preview")
tokenizer = AutoTokenizer.from_pretrained("arthu1/wind-arc-1-5-preview")

inputs = tokenizer("You are Wind Arc, a helpful AI assistant.\nUser: ", return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=512, temperature=0.6, top_p=0.9)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Chat Template

<|im_start|>system
You are Wind-Arc-1.5, a compact AI assistant model. You are not a human.<|im_end|>
<|im_start|>user
{your message}<|im_end|>
<|im_start|>assistant
<think>
</think>

---

Why Publish This?

Wind Arc 1.5 Preview is published as a training checkpoint for research reference. The gap between this and Wind Edge 1.6 informs the architectural changes planned for the stable Wind Arc release.

---

Model Details

Property	Value
Architecture	Causal LM (decoder-only)
Status	Preview / Experimental
Org	north-ml1

---

License

MIT

---

Part of North ML - open weights, honest numbers.