issdandavis/polly-scbe-7b-v2-merged

Polly SCBE 7B v2 — Merged

Polly is a conversational + governance-aware language model fine-tuned on the SCBE-AETHERMOORE framework. Named after a raven familiar from 528 pages of Everweave RPG lore — the same corpus that seeded the Sacred Tongues tokenizer.

This is the merged (full-weight) version. LoRA adapter at issdandavis/polly-scbe-7b-v2.

Architecture

• Base: Qwen/Qwen2.5-7B-Instruct
• Fine-tuning: QLoRA (4-bit bitsandbytes), SFT
• Training: Google Colab (T4/A100)
• Method: merge_and_unload() — LoRA weights baked into base

What is SCBE / Sacred Tongues?

SCBE-AETHERMOORE is a 14-layer AI safety pipeline using hyperbolic geometry (Poincare ball model). The Sacred Tongues tokenizer encodes text through 6 language-archetype dimensions:

Code	Tongue	Spirit	Role
KO	Kor'aelin	Python / precise	User intent
AV	Avali	JS / reactive	System context
RU	Runethic	Rust / ethical	Structural reasoning
CA	Cassisivadan	Mathematica / symbolic	Model output
UM	Umbroth	Haskell / pure-functional	Deep abstraction
DR	Draumric	Markdown / narrative	Integrity seal

Unlike standard tokenizers built from frequency statistics, Sacred Tongues was seeded from 528 pages of collaborative RPG world-building — giving the vocabulary semantic density from character arcs, dimensional magic, and governance rituals rather than raw token frequency.

Chat Template

The tokenizer routes each conversation role through its canonical tongue:

• system → Avali (header context)
• user → Kor'aelin (intent)
• assistant → Cassisivadan (output)
• tool → Draumric (integrity)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "issdandavis/polly-scbe-7b-v2-merged",
    device_map="auto",
    load_in_4bit=True,
)
tokenizer = AutoTokenizer.from_pretrained("issdandavis/polly-scbe-7b-v2-merged")

messages = [{"role": "user", "content": "Explain the SCBE governance pipeline."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Related Models

Model	Base	Notes
polly-scbe-7b-v2	Qwen2.5-7B	LoRA adapter (this run)
polly-prime-r7-qwen-1.5b	Qwen2.5-Coder-1.5B	r7 coder fine-tune
polly-r8-qwen-0.5b	Qwen2.5-Coder-0.5B	r8 Kaggle run

Project

• GitHub: SCBE-AETHERMOORE
• Developer: Issac Davis — solo builder, Port Angeles WA