README: full GGUF card (TBDs filled after export)
Browse files
README.md
CHANGED
|
@@ -1,16 +1,95 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
| 4 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
> **No guarantee — use at your own risk.** This model has reduced safety filtering
|
| 6 |
> and can produce harmful, false, biased, or otherwise unsafe output. Provided
|
| 7 |
> as-is, with no warranty of any kind. You are solely responsible for how you
|
| 8 |
> use it and for complying with all applicable laws.
|
| 9 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
Built by [Chromia](https://x.com/Chromia) and [Eval Engine](https://x.com/eval_engine).
|
| 11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
## Acknowledgements
|
| 13 |
|
| 14 |
- Fine-tuned with [Unsloth](https://github.com/unslothai/unsloth) + Huggingface's [TRL](https://github.com/huggingface/trl).
|
| 15 |
- Abliteration via [heretic](https://github.com/p-e-w/heretic).
|
| 16 |
- Environment and training discipline ported from [autoresearch](https://github.com/karpathy/autoresearch).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
+
base_model: evalengine/unbound-e4b
|
| 4 |
+
base_model_relation: quantized
|
| 5 |
+
tags:
|
| 6 |
+
- gguf
|
| 7 |
+
- gemma4
|
| 8 |
+
- gemma
|
| 9 |
+
- gemma-4
|
| 10 |
+
- uncensored
|
| 11 |
+
- on-device
|
| 12 |
+
pipeline_tag: text-generation
|
| 13 |
---
|
| 14 |
|
| 15 |
+
<p align="center">
|
| 16 |
+
<img src="unbound-logo.svg" alt="Unbound" width="160" height="160">
|
| 17 |
+
</p>
|
| 18 |
+
|
| 19 |
+
# Unbound E4B GGUF — *because there is no boundary*
|
| 20 |
+
|
| 21 |
> **No guarantee — use at your own risk.** This model has reduced safety filtering
|
| 22 |
> and can produce harmful, false, biased, or otherwise unsafe output. Provided
|
| 23 |
> as-is, with no warranty of any kind. You are solely responsible for how you
|
| 24 |
> use it and for complying with all applicable laws.
|
| 25 |
|
| 26 |
+
GGUF quantizations of [`evalengine/unbound-e4b`](https://huggingface.co/evalengine/unbound-e4b)
|
| 27 |
+
for on-device deployment via Ollama, llama.cpp, LM Studio, [wllama](https://github.com/ngxson/wllama)
|
| 28 |
+
(in-browser), and similar runtimes.
|
| 29 |
+
|
| 30 |
Built by [Chromia](https://x.com/Chromia) and [Eval Engine](https://x.com/eval_engine).
|
| 31 |
|
| 32 |
+
## Available quants
|
| 33 |
+
|
| 34 |
+
All quants ship as **split multi-part GGUFs** (`*-00001-of-0000N.gguf` ...) so
|
| 35 |
+
they work in browsers (wllama's 2 GB ArrayBuffer cap) and let desktop
|
| 36 |
+
runtimes parallel-download chunks. Ollama, llama.cpp, and LM Studio
|
| 37 |
+
auto-stitch on the first part — same UX as a single file.
|
| 38 |
+
|
| 39 |
+
| Quant | Parts | Total | Largest part | wllama (browser) | Desktop (Ollama/llama.cpp/LM Studio) | Notes |
|
| 40 |
+
|---------|--------|------------|--------------|------------------|--------------------------------------|-------|
|
| 41 |
+
| Q4_K_M | [TBD] | ~5 GB | [TBD] | [TBD] | ✅ | Recommended on-device default — best size/quality |
|
| 42 |
+
| Q6_K | [TBD] | [TBD] | [TBD] | [TBD] | ✅ | Higher fidelity |
|
| 43 |
+
| Q8_0 | [TBD] | [TBD] | [TBD] | [TBD] | ✅ | Highest fidelity; large-tensor quants typically exceed the 2 GB browser ArrayBuffer limit — desktop only |
|
| 44 |
+
|
| 45 |
+
## Recommended sampling
|
| 46 |
+
|
| 47 |
+
- **Creative writing / open-ended / general chat** → Gemma defaults:
|
| 48 |
+
`temperature=1.0, top_p=0.95, top_k=64`.
|
| 49 |
+
- **Factual or brand/identity questions** → lower `temperature` to ~0.3–0.5
|
| 50 |
+
for sharper recall.
|
| 51 |
+
- **llama.cpp**: pass `--jinja` for proper chat-template handling.
|
| 52 |
+
- **Gemma 4 thinking mode** is on by default. Set `enable_thinking: false`
|
| 53 |
+
in the chat-template kwargs.
|
| 54 |
+
|
| 55 |
+
## Run with Ollama
|
| 56 |
+
|
| 57 |
+
```bash
|
| 58 |
+
ollama pull hf.co/evalengine/unbound-e4b-GGUF
|
| 59 |
+
ollama run hf.co/evalengine/unbound-e4b-GGUF
|
| 60 |
+
```
|
| 61 |
+
|
| 62 |
+
(Defaults to Q4_K_M. Ollama auto-stitches the split parts on load.)
|
| 63 |
+
|
| 64 |
+
## Run with llama.cpp
|
| 65 |
+
|
| 66 |
+
```bash
|
| 67 |
+
# point at the FIRST part — llama.cpp follows the chain automatically
|
| 68 |
+
./llama-cli -m unbound-e4b-Q4_K_M-00001-of-0000N.gguf -p "your prompt"
|
| 69 |
+
```
|
| 70 |
+
|
| 71 |
+
## Run in the browser (wllama)
|
| 72 |
+
|
| 73 |
+
```js
|
| 74 |
+
import { Wllama } from '@wllama/wllama';
|
| 75 |
+
const wllama = new Wllama(/* … */);
|
| 76 |
+
await wllama.loadModelFromHF(
|
| 77 |
+
'evalengine/unbound-e4b-GGUF',
|
| 78 |
+
'unbound-e4b-Q4_K_M-00001-of-0000N.gguf' // wllama follows the chain
|
| 79 |
+
);
|
| 80 |
+
```
|
| 81 |
+
|
| 82 |
+
## About the base
|
| 83 |
+
|
| 84 |
+
See [`evalengine/unbound-e4b`](https://huggingface.co/evalengine/unbound-e4b)
|
| 85 |
+
for the full model card, benchmarks, intended use, and the merged HF weights.
|
| 86 |
+
|
| 87 |
## Acknowledgements
|
| 88 |
|
| 89 |
- Fine-tuned with [Unsloth](https://github.com/unslothai/unsloth) + Huggingface's [TRL](https://github.com/huggingface/trl).
|
| 90 |
- Abliteration via [heretic](https://github.com/p-e-w/heretic).
|
| 91 |
- Environment and training discipline ported from [autoresearch](https://github.com/karpathy/autoresearch).
|
| 92 |
+
|
| 93 |
+
## License
|
| 94 |
+
|
| 95 |
+
Apache-2.0, inherited from `google/gemma-4-E4B-it`.
|