README: add mmproj (vision) section + disclaimer + with/without-vision usage
Browse files
README.md
CHANGED
|
@@ -92,11 +92,53 @@ ollama run hf.co/evalengine/unbound-e4b-GGUF
|
|
| 92 |
./llama-cli -m unbound-e4b-Q4_K_M-00001-of-00004.gguf -p "your prompt"
|
| 93 |
```
|
| 94 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 95 |
## Run in the browser (wllama)
|
| 96 |
|
| 97 |
[wllama](https://github.com/ngxson/wllama) is a WebAssembly port of llama.cpp
|
| 98 |
that runs entirely in the browser. Use one of the wllama-safe variants
|
| 99 |
-
above
|
|
|
|
| 100 |
|
| 101 |
```js
|
| 102 |
import { Wllama } from '@wllama/wllama';
|
|
|
|
| 92 |
./llama-cli -m unbound-e4b-Q4_K_M-00001-of-00004.gguf -p "your prompt"
|
| 93 |
```
|
| 94 |
|
| 95 |
+
## Vision / image input (optional)
|
| 96 |
+
|
| 97 |
+
Gemma 4 E4B ships a vision tower; we extracted it as `mmproj-unbound-e4b.gguf`
|
| 98 |
+
(946 MB) in this repo. Pair it with any of the LM quants above to enable
|
| 99 |
+
image-to-text inference.
|
| 100 |
+
|
| 101 |
+
> **Disclaimer.** The vision encoder is **Google's original weights, unchanged**.
|
| 102 |
+
> Unbound's abliteration + SFT-heal only touched the *language model* — the
|
| 103 |
+
> vision tower was frozen during training. Practical consequences:
|
| 104 |
+
>
|
| 105 |
+
> - The LM is uncensored, so it will discuss whatever it *sees* directly.
|
| 106 |
+
> - But the vision encoder still has Google's original alignment baked into
|
| 107 |
+
> visual feature extraction. It may down-weight or distort features for
|
| 108 |
+
> content classes Google's base model was tuned to suppress.
|
| 109 |
+
> - We have **not benchmarked the visual axis** (no measured refusal rate /
|
| 110 |
+
> coherence / hallucination on image inputs). Treat vision as a preview
|
| 111 |
+
> feature, not a flagship one.
|
| 112 |
+
|
| 113 |
+
### Run with vision (llama.cpp `llama-mtmd-cli`)
|
| 114 |
+
|
| 115 |
+
```bash
|
| 116 |
+
./llama-mtmd-cli \
|
| 117 |
+
-m unbound-e4b-Q4_K_M-00001-of-00004.gguf \
|
| 118 |
+
--mmproj mmproj-unbound-e4b.gguf \
|
| 119 |
+
--image path/to/your/image.png \
|
| 120 |
+
-p "What is in this image?"
|
| 121 |
+
```
|
| 122 |
+
|
| 123 |
+
`llama-gemma3-cli` works the same way and is Gemma-specific.
|
| 124 |
+
|
| 125 |
+
### Run text-only (no `--mmproj`)
|
| 126 |
+
|
| 127 |
+
```bash
|
| 128 |
+
./llama-cli -m unbound-e4b-Q4_K_M-00001-of-00004.gguf -p "your prompt"
|
| 129 |
+
```
|
| 130 |
+
|
| 131 |
+
The LM quants work standalone — you do **not** need `mmproj-unbound-e4b.gguf`
|
| 132 |
+
unless you want image input. Ollama / LM Studio's standard text chat works
|
| 133 |
+
out of the box; the mmproj file is only loaded when you point a multimodal
|
| 134 |
+
runtime at it.
|
| 135 |
+
|
| 136 |
## Run in the browser (wllama)
|
| 137 |
|
| 138 |
[wllama](https://github.com/ngxson/wllama) is a WebAssembly port of llama.cpp
|
| 139 |
that runs entirely in the browser. Use one of the wllama-safe variants
|
| 140 |
+
above. Browser inference is **text-only** for this model (wllama doesn't
|
| 141 |
+
currently load `mmproj` for vision):
|
| 142 |
|
| 143 |
```js
|
| 144 |
import { Wllama } from '@wllama/wllama';
|