구성 / Components
- Main model: Qwen/Qwen3.5-2B.
- Quantized base and F16 visual projector: unmodified files from
unsloth/Qwen3.5-2B-GGUF.
- Trained component: Office LoRA v1, strength 1.0, last full-attention block 23.
- Included decision companion:
d1-2B-Q4_K_M.gguf, an unchanged copy from
LiquidAI/d1-2B-GGUF.
The pretrained companion has its own LFM Open License v1.0.
All four runtime files are included so a download contains the tested setup.
- Visual encoder/projector and decision companion weights were not trained by Narae.
- The application routes components when needed. This is not a jointly trained sparse MoE.
실제 학습 / Actual training
Table with columns: Item, Recorded result| Item | Recorded result |
|---|
| Synthetic training / validation dialogues | 32 / 3 |
| Epochs / optimizer updates | 2 / 16 |
| Trainable parameters / changed tensors | 638,976 / 14 |
| LoRA rank / alpha | 16 / 32 |
| Validation loss before / after | 1.570969 / 1.550289 |
| Fixed held-out next-action score before / after | 7/8 / 7/8 |
The held-out accuracy did not improve. Later v2 candidates were not adopted.
The GGUF adapter's 14 tensors were checked bit-for-bit against the saved trained
adapter. Frozen BF16 training versus Q4_K_M inference is a quantization difference.
Training scripts and synthetic dataset hashes are in the
reproduction folder.
사용 / Use
For one-click installation of all four files, use Narae Hangul Office
preview.25 or later. The preview.25 application source
includes the pinned download catalog. Select 나래4 in the local AI settings
and install its verified files. New installations select Narae4 by default;
existing explicit AI choices are preserved.
The application verifies SHA256, the loaded model alias and the actual LoRA.
Tool execution, Hangul/Windows integration and authenticated school UI access
require the application and the user's own local setup; weights alone do not
provide those tools or authorization.
For llama.cpp, download the GGUF base and adapter and always load both:
llama-server --model Qwen3.5-2B-Q4_K_M.gguf --lora Narae4-Qwen3.5-2B-Office-LoRA-v1-f32.gguf --alias Narae4-Qwen3.5-2B-Office-LoRA-v1 --host 127.0.0.1 --ctx-size 8192 --gpu-layers 0 --jinja
Add --mmproj mmproj-F16.gguf --no-mmproj-offload for image input.
Use a llama.cpp build supporting Qwen3.5 and its visual projector. The PEFT
adapter pins the original Qwen base revision in adapter_config.json; the full
Transformers base is not included. See the
runtime and workflow guide.
실행 검증과 한계 / Evaluation and limitations
On 2026-11-01 the v1 model with the preview.24 runtime completed one explicitly
authorized synthetic 1-won requisition cycle: enter, choose an observed budget,
enter an item, attach original bytes, save, reopen from a server query, verify
attachment bytes/SHA256, delete, and verify absence in a fresh query. It took
20 tool stages and 131.296 seconds; the test record was deleted. Approval
submission and actual purchasing/payment were not performed. This measures
the model plus bounded runtime/controller, not free-form model-only autonomy.
Earlier synthetic document/vision smoke tests and the above single integration
run do not establish general school-work accuracy. Missing budget choices,
ambiguous document identities and uncertain writes stop the workflow. Current
financial approval submission, arbitrary existing-record deletion, multiple
budgets and complete virtualized-grid traversal are not validated capabilities.
See evaluation-summary.json for the distinct training and runtime evidence.
No real school documents, student/staff identities, portal screenshots, sessions,
credentials, private conversations or live execution logs are included.
파일 / Files
Table with columns: File, Bytes, SHA256| File | Bytes | SHA256 |
|---|
| Qwen3.5-2B-Q4_K_M.gguf | 1280835840 | aaf42c8b7c3cab2bf3d69c355048d4a0ee9973d48f16c731c0520ee914699223 |
| Narae4-Qwen3.5-2B-Office-LoRA-v1-f32.gguf | 2557408 | 19f20486e744bcb9636efcdaf6583c1b9bef545733987a8837e63afdfeecf105 |
| mmproj-F16.gguf | 668227264 | 7035e9cb8d7c6a9681d07eef9a364783e86ea4cd73faab2eabb4f43a101830c7 |
| d1-2B-Q4_K_M.gguf | 1099629888 | adca5789ad30ae6027e8e936749bb9f19673097492ab26fa1c5b69643ca50b4b |
| adapter_model.safetensors | 2557984 | 3380c0dfbee1737fa7f741d5ec475efdbfbba15206e78373a93592f4c745fc4d |
| adapter_config.json | 1253 | 70151bec3a55e5b846976b58a3f778427716947fe3e2001deec569a5a762c4c7 |
라이선스 / License
The Qwen-derived files and Narae LoRA use Apache-2.0;
retain LICENSE and NOTICE. The included LiquidAI companion
uses LFM Open License v1.0 in D1-LICENSE.
See COMPONENT_LICENSES.md for the file-specific terms.
The application's MIT license does not replace the model licenses.