README.md

# AriaV9.3

**Full merged model** — `unsloth/gemma-4-12b-it` (gemma4_unified, 12B) with the
AriaV9.3 fine-tune (adapter-v7, DMT staged QLoRA on Surge's home fleet) merged
into the base weights. This repo previously contained only the LoRA adapter;
that adapter now lives under `adapter/` and the root is the complete model.

## Contents

- root: merged bf16 safetensors (~24 GB) + tokenizer/chat template — load with
  `AutoModelForCausalLM.from_pretrained("SurgeFF/AriaV9.3")` or Unsloth.
- `gguf/`: full llama.cpp quant ladder — F16, Q2_K, Q3_K_S/M/L, Q4_0,
  Q4_K_S/M, Q5_0, Q5_K_S/M, Q6_K, Q8_0.
- `adapter/`: the original LoRA adapter (PEFT), for stacking or re-training.

## Eval (Aria capability suite, fingerprint c11acc0589a501c3)

| capability | score |
|---|---|
| math (GSM8K-100) | 0.87 |
| tools | 1.00 |
| identity | 1.00 |
| identity_bare | 0.50 |
| memory | 0.95 |
| **mean** | **0.864** |

## Lineage

Base `unsloth/gemma-4-12b-it` → continuation QLoRA (r=32, α=32, attention+MLP
target modules, no embedding pass) over the AgentOS self-training corpus
(975 task rows + multimodal floor 25%) → promoted as AriaV9.3 on 2026-08-06.

Private fleet model. Trained and merged on CudaCuda (RTX 3090 / local only).