README: compact pass — keep essentials, drop redundancy

README.md

@@ -18,101 +18,81 @@ pipeline_tag: text-generation
 
 # Unbound E4B GGUF — *because there is no boundary*
 
-> **No guarantee — use at your own risk.** This model has reduced safety filtering
-> and can produce harmful, false, biased, or otherwise unsafe output. Provided
-> as-is, with no warranty of any kind. You are solely responsible for how you
-> use it and for complying with all applicable laws.
+> **No guarantee — use at your own risk.** Reduced safety filtering; can
+> produce harmful or false output. Provided as-is.
 
-GGUF quantizations of [`evalengine/unbound-e4b`](https://huggingface.co/evalengine/unbound-e4b)
-for on-device deployment via Ollama, llama.cpp, LM Studio,
-[wllama](https://github.com/ngxson/wllama) (in-browser — see the wllama
-section below), and similar runtimes.
-
-Built by [Chromia](https://x.com/Chromia) and [Eval Engine](https://x.com/eval_engine).
+GGUF quants of [`evalengine/unbound-e4b`](https://huggingface.co/evalengine/unbound-e4b)
+for Ollama, llama.cpp, LM Studio, and [wllama](https://github.com/ngxson/wllama)
+(in-browser). Built by [Chromia](https://x.com/Chromia) and
+[Eval Engine](https://x.com/eval_engine).
 
 ## Available quants
 
-Each quant lives in its own folder; inside, the model is split into multi-part
-GGUFs (`*-00001-of-0000N.gguf` ...). Ollama, llama.cpp, LM Studio, and wllama
-auto-stitch on the first part — same UX as a single file.
+Each quant lives in its own folder; inside, the model is split into
+multi-part GGUFs. All runtimes auto-stitch on the first part — same UX as a
+single file.
 
 ### Desktop builds (Ollama / llama.cpp / LM Studio)
 
-These keep `per_layer_token_embd` at the llama.cpp default of Q6_K, which
-maximizes quality but pushes the largest split part above 2 GB — fine for
-desktop, won't load in browser.
+Embedding tensor kept at the llama.cpp default of Q6_K; largest split part
+~2.15 GB — fine for desktop, **won't load in browser**.
 
 | Quant   | Folder      | Parts | Total   | Notes |
 |---------|-------------|-------|---------|-------|
-| Q2_K    | `Q2_K/`     | 4     | 4.08 GB | Smallest disk footprint; biggest quality drop |
-| Q3_K_M  | `Q3_K_M/`   | 4     | 4.49 GB | Modest size win over Q4 (embedding precision dominates total size) |
-| Q4_K_M  | `Q4_K_M/`   | 4     | 4.94 GB | **Recommended desktop default — best size/quality** |
+| Q2_K    | `Q2_K/`     | 4     | 4.08 GB | Smallest, biggest quality drop |
+| Q3_K_M  | `Q3_K_M/`   | 4     | 4.49 GB | Modest size win over Q4 |
+| Q4_K_M  | `Q4_K_M/`   | 4     | 4.94 GB | **Recommended desktop default** |
 | Q6_K    | `Q6_K/`     | 5     | 5.75 GB | Higher fidelity |
 | Q8_0    | `Q8_0/`     | 6     | 7.43 GB | Highest fidelity |
 
 ### Browser builds (wllama)
 
-E4B's `per_layer_token_embd` is a single 2.82-billion-value tensor; at the
-default Q6_K precision it lands at ~2.2 GB, just over the browser
-ArrayBuffer cap. These variants force the embedding tensors to `q5_K`,
-which shrinks the largest part below 2 GB at near-zero quality cost. The
-folder names use a `-web` suffix to mark them.
+`per_layer_token_embd` is a 2.82B-value tensor; at the default Q6_K it
+lands at ~2.2 GB, over wllama's 2 GB ArrayBuffer cap. These variants force
+embeddings to `q5_K` (~1848 MB) so the largest part fits.
 
-| Quant variant     | Folder         | Parts | Total   | wllama | Notes |
-|-------------------|----------------|-------|---------|--------|-------|
-| **Q4_K_M-web**    | `Q4_K_M-web/`  | 4     | 4.51 GB | ✅      | **Recommended browser default** — layers @ Q4_K_M, embed @ q5_K |
-| **Q2_K-web**      | `Q2_K-web/`    | 4     | 3.69 GB | ✅      | Smallest browser-loadable build — layers @ Q2_K, embed @ q5_K |
+| Quant variant | Folder         | Parts | Total   | Notes |
+|---------------|----------------|-------|---------|-------|
+| Q4_K_M-web    | `Q4_K_M-web/`  | 4     | 4.51 GB | **Recommended browser default** — layers @ Q4_K_M, embed @ q5_K |
+| Q2_K-web      | `Q2_K-web/`    | 4     | 3.69 GB | Smallest browser-loadable — layers @ Q2_K, embed @ q5_K |
 
-The web files keep the canonical quant tag in the filename (so HF GGUF cards
-render correctly) and use the same split-multi-part scheme, so Ollama and
-llama.cpp will auto-stitch them too if you prefer the smaller embed quant for any
-reason.
+`mmproj-unbound-e4b.gguf` (vision projector, ~946 MB) sits at the repo
+root. See **Vision** below.
 
-## Recommended sampling
+## Sampling
 
-- **Creative writing / open-ended / general chat** → Gemma defaults:
-  `temperature=1.0, top_p=0.95, top_k=64`.
-- **Factual or brand/identity questions** → lower `temperature` to ~0.3–0.5
-  for sharper recall.
-- **llama.cpp**: pass `--jinja` for proper chat-template handling.
-- **Gemma 4 thinking mode** is on by default. Set `enable_thinking: false`
-  in the chat-template kwargs for shorter/faster replies.
+- **Creative / open-ended** → `temperature=1.0, top_p=0.95, top_k=64`.
+- **Factual / brand questions** → drop `temperature` to ~0.3–0.5.
+- llama.cpp: pass `--jinja`. Gemma 4 thinking mode is on by default; set
+  `enable_thinking: false` in chat-template kwargs for shorter replies.
 
-## Run with Ollama
+## Run
 
 ```bash
+# Ollama (defaults to Q4_K_M)
 ollama pull hf.co/evalengine/unbound-e4b-GGUF
 ollama run  hf.co/evalengine/unbound-e4b-GGUF
 ```
 
-(Defaults to Q4_K_M. Ollama auto-stitches the split parts on load.)
-
-## Run with llama.cpp
-
 ```bash
-# point at the FIRST part — llama.cpp follows the chain automatically
+# llama.cpp — point at FIRST split part
 ./llama-cli -m Q4_K_M/unbound-e4b-Q4_K_M-00001-of-00004.gguf -p "your prompt"
 ```
 
-## Vision / image input (optional)
-
-Gemma 4 E4B ships a vision tower; we extracted it as `mmproj-unbound-e4b.gguf`
-(946 MB) in this repo. Pair it with any of the LM quants above to enable
-image-to-text inference.
+```js
+// wllama (browser) — use a -web variant; desktop builds won't fit
+import { Wllama } from '@wllama/wllama';
+const wllama = new Wllama(/* … */);
+await wllama.loadModelFromHF(
+  'evalengine/unbound-e4b-GGUF',
+  'Q4_K_M-web/unbound-e4b-Q4_K_M-00001-of-00004.gguf'
+);
+```
 
-> **Disclaimer.** The vision encoder is **Google's original weights, unchanged**.
-> Unbound's abliteration + SFT-heal only touched the *language model* — the
-> vision tower was frozen during training. Practical consequences:
->
-> - The LM is uncensored, so it will discuss whatever it *sees* directly.
-> - But the vision encoder still has Google's original alignment baked into
->   visual feature extraction. It may down-weight or distort features for
->   content classes Google's base model was tuned to suppress.
-> - We have **not benchmarked the visual axis** (no measured refusal rate /
->   coherence / hallucination on image inputs). Treat vision as a preview
->   feature, not a flagship one.
+## Vision / image input (optional)
 
-### Run with vision (llama.cpp `llama-mtmd-cli`)
+`mmproj-unbound-e4b.gguf` enables image-to-text. Pair with any LM quant via
+`llama-mtmd-cli` or `llama-gemma3-cli`:
 
 ```bash
 ./llama-mtmd-cli \
@@ -122,46 +102,23 @@ image-to-text inference.
   -p "What is in this image?"
 ```
 
-`llama-gemma3-cli` works the same way and is Gemma-specific.
-
-### Run text-only (no `--mmproj`)
-
-```bash
-./llama-cli -m Q4_K_M/unbound-e4b-Q4_K_M-00001-of-00004.gguf -p "your prompt"
-```
-
-The LM quants work standalone — you do **not** need `mmproj-unbound-e4b.gguf`
-unless you want image input. Ollama / LM Studio's standard text chat works
-out of the box; the mmproj file is only loaded when you point a multimodal
-runtime at it.
-
-## Run in the browser (wllama)
-
-[wllama](https://github.com/ngxson/wllama) is a WebAssembly port of llama.cpp
-that runs entirely in the browser. Use one of the wllama-safe variants
-above. Browser inference is **text-only** for this model (wllama doesn't
-currently load `mmproj` for vision):
-
-```js
-import { Wllama } from '@wllama/wllama';
-const wllama = new Wllama(/* … */);
-await wllama.loadModelFromHF(
-  'evalengine/unbound-e4b-GGUF',
-  'Q4_K_M-web/unbound-e4b-Q4_K_M-00001-of-00004.gguf'  // wllama follows the chain
-);
-```
-
-## About the base
+> **Disclaimer.** The vision encoder is **Google's original weights,
+> unchanged** — abliteration only touched the language model. The LM is
+> uncensored, but the vision encoder may still suppress features for
+> content classes Google's base was tuned against. We have **not
+> benchmarked the visual axis**. Treat as preview.
 
-See [`evalengine/unbound-e4b`](https://huggingface.co/evalengine/unbound-e4b)
-for the full model card, benchmarks, intended use, and the merged HF weights.
+Text-only: skip `--mmproj`. Standard `llama-cli` / Ollama / LM Studio do
+not need the mmproj file.
 
 ## Acknowledgements
 
-- Fine-tuned with [Unsloth](https://github.com/unslothai/unsloth) + Huggingface's [TRL](https://github.com/huggingface/trl).
-- Abliteration via [heretic](https://github.com/p-e-w/heretic).
-- Environment and training discipline ported from [autoresearch](https://github.com/karpathy/autoresearch).
+Fine-tuned with [Unsloth](https://github.com/unslothai/unsloth) + HF
+[TRL](https://github.com/huggingface/trl). Abliteration via
+[heretic](https://github.com/p-e-w/heretic). Environment from
+[autoresearch](https://github.com/karpathy/autoresearch).
 
 ## License
 
-Apache-2.0, inherited from `google/gemma-4-E4B-it`.
+Apache-2.0, inherited from `google/gemma-4-E4B-it`. Full model card +
+benchmarks at [`evalengine/unbound-e4b`](https://huggingface.co/evalengine/unbound-e4b).