Add Doppel model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: coreml
|
| 4 |
+
pipeline_tag: text-generation
|
| 5 |
+
base_model: z-lab/Qwen3.5-9B-DFlash
|
| 6 |
+
inference: false
|
| 7 |
+
tags:
|
| 8 |
+
- coreml
|
| 9 |
+
- apple-silicon
|
| 10 |
+
- macos
|
| 11 |
+
- neural-engine
|
| 12 |
+
- speculative-decoding
|
| 13 |
+
- dflash
|
| 14 |
+
- doppel
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Doppel speculative-decoding companion
|
| 18 |
+
|
| 19 |
+
This repository is Doppel's distribution channel for an app-specific Core ML
|
| 20 |
+
conversion of
|
| 21 |
+
[`z-lab/Qwen3.5-9B-DFlash`](https://huggingface.co/z-lab/Qwen3.5-9B-DFlash).
|
| 22 |
+
Doppel uses it as a Neural Engine draft companion while its target model
|
| 23 |
+
performs verification.
|
| 24 |
+
|
| 25 |
+
## Intended use
|
| 26 |
+
|
| 27 |
+
These artifacts are intended to be downloaded, verified, and invoked by
|
| 28 |
+
compatible Doppel releases on Apple-silicon Macs running macOS 15 or later.
|
| 29 |
+
They are an optimization for Doppel's on-device inference path.
|
| 30 |
+
|
| 31 |
+
This is not a standalone language model, text generator, or supported
|
| 32 |
+
developer API. It omits a standalone vocabulary head and depends on Doppel's
|
| 33 |
+
exact target model, hidden-state contract, cache lineage, verification logic,
|
| 34 |
+
and runtime orchestration. Loading the Core ML packages individually does not
|
| 35 |
+
constitute a valid or supported inference path. For research or general DFlash
|
| 36 |
+
use, start with the upstream model and project.
|
| 37 |
+
|
| 38 |
+
## Artifact and integrity notes
|
| 39 |
+
|
| 40 |
+
The source checkpoint is split into app-specific Core ML packages. The
|
| 41 |
+
repository's `manifest.json` is the authoritative compatibility and artifact
|
| 42 |
+
inventory; it records the integrity hashes checked by Doppel. Core ML's local
|
| 43 |
+
compilation cache is deliberately not distributed.
|
| 44 |
+
|
| 45 |
+
The repository contains no user prompts, conversations, model caches,
|
| 46 |
+
telemetry, or other customer data.
|
| 47 |
+
|
| 48 |
+
## Limitations
|
| 49 |
+
|
| 50 |
+
This companion is qualified only as part of Doppel's complete speculative
|
| 51 |
+
decoder. It must not be evaluated as an independent generator, and compatibility
|
| 52 |
+
must not be inferred from a shared tokenizer or architecture name.
|
| 53 |
+
|
| 54 |
+
## License and attribution
|
| 55 |
+
|
| 56 |
+
The DFlash checkpoint is provided under the
|
| 57 |
+
[Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).
|
| 58 |
+
Doppel pairs the conversion with the separately licensed
|
| 59 |
+
[`Ornith-1.0-9B`](https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B)
|
| 60 |
+
target. This converted distribution remains subject to the applicable
|
| 61 |
+
upstream terms. Doppel provides no support or warranty for use outside the
|
| 62 |
+
Doppel app.
|