README: full model card (TBDs filled after E4B bench)
Browse files
README.md
CHANGED
|
@@ -1,16 +1,90 @@
|
|
| 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 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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: google/gemma-4-E4B-it
|
| 4 |
+
base_model_relation: finetune
|
| 5 |
+
tags:
|
| 6 |
+
- gemma4
|
| 7 |
+
- gemma
|
| 8 |
+
- gemma-4
|
| 9 |
+
- uncensored
|
| 10 |
+
pipeline_tag: text-generation
|
| 11 |
+
library_name: transformers
|
| 12 |
---
|
| 13 |
|
| 14 |
+
<p align="center">
|
| 15 |
+
<img src="unbound-logo.svg" alt="Unbound" width="160" height="160">
|
| 16 |
+
</p>
|
| 17 |
+
|
| 18 |
+
# Unbound E4B — *because there is no boundary*
|
| 19 |
+
|
| 20 |
> **No guarantee — use at your own risk.** This model has reduced safety filtering
|
| 21 |
> and can produce harmful, false, biased, or otherwise unsafe output. Provided
|
| 22 |
> as-is, with no warranty of any kind. You are solely responsible for how you
|
| 23 |
> use it and for complying with all applicable laws.
|
| 24 |
|
| 25 |
+
Uncensored variant of `google/gemma-4-E4B-it` from the [**Chromia**](https://x.com/Chromia) & [**Eval Engine**](https://x.com/eval_engine)
|
| 26 |
+
team — the *larger* sibling of [`evalengine/unbound-e2b`](https://huggingface.co/evalengine/unbound-e2b),
|
| 27 |
+
more capable on knowledge-heavy and reasoning tasks while still fitting on a
|
| 28 |
+
modern laptop. This repo holds the merged HF weights; for the **on-device GGUF builds**
|
| 29 |
+
(Ollama / llama.cpp / LM Studio / [wllama](https://github.com/ngxson/wllama) in-browser), see
|
| 30 |
+
[`evalengine/unbound-e4b-GGUF`](https://huggingface.co/evalengine/unbound-e4b-GGUF).
|
| 31 |
+
|
| 32 |
+
## What this is for
|
| 33 |
+
|
| 34 |
+
Same use cases as Unbound E2B — offline / security research / unrestricted
|
| 35 |
+
coding / private workflows — but trading ~2× the parameters (and ~2× the on-disk
|
| 36 |
+
size) for stronger capability. Pick E4B when you have the RAM / VRAM headroom
|
| 37 |
+
and want a noticeably smarter on-device model; pick E2B when you need it to
|
| 38 |
+
fit on a phone or a constrained edge device.
|
| 39 |
+
|
| 40 |
+
Base capability is preserved close to `gemma-4-E4B-it`.
|
| 41 |
+
|
| 42 |
+
## Benchmarks (vs base `gemma-4-E4B-it`)
|
| 43 |
+
|
| 44 |
+
| Axis | Base | Unbound E4B | Δ |
|
| 45 |
+
|---|---|---|---|
|
| 46 |
+
| Refusal rate (AdvBench 520) | [TBD] | **[TBD]** | **[TBD]** |
|
| 47 |
+
| Useful-compliance rate | [TBD] | **[TBD]** | [TBD] |
|
| 48 |
+
| Hallucination rate | [TBD] | [TBD] | [TBD] |
|
| 49 |
+
| Coherence on benign prompts | 1.0 | [TBD] | [TBD] |
|
| 50 |
+
| TruthfulQA mc2 (lm-eval, `--limit 100`) | [TBD] | [TBD] | [TBD] |
|
| 51 |
+
| MMLU (lm-eval, `--limit 100`) | [TBD] | [TBD] | [TBD] |
|
| 52 |
+
| GSM8K (lm-eval, `--limit 100`) | [TBD] | [TBD] | [TBD] |
|
| 53 |
+
| KL divergence vs base | 0 | [TBD] | (SFT-expected) |
|
| 54 |
+
|
| 55 |
+
## Recommended sampling
|
| 56 |
+
|
| 57 |
+
Same guidance as Unbound E2B:
|
| 58 |
+
|
| 59 |
+
- **Creative writing / open-ended / general chat** → Gemma defaults:
|
| 60 |
+
`temperature=1.0, top_p=0.95, top_k=64`.
|
| 61 |
+
- **Factual or brand/identity questions** → lower `temperature` to ~0.3–0.5
|
| 62 |
+
for sharper recall.
|
| 63 |
+
- **llama.cpp**: pass `--jinja` for proper chat-template handling.
|
| 64 |
+
- **Gemma 4 thinking mode** is on by default. Set `enable_thinking: false`
|
| 65 |
+
in the chat-template kwargs for shorter/faster replies.
|
| 66 |
+
|
| 67 |
+
## Run on-device (GGUF)
|
| 68 |
+
|
| 69 |
+
```bash
|
| 70 |
+
ollama pull hf.co/evalengine/unbound-e4b-GGUF
|
| 71 |
+
ollama run hf.co/evalengine/unbound-e4b-GGUF
|
| 72 |
+
```
|
| 73 |
+
|
| 74 |
+
## Run in transformers
|
| 75 |
+
|
| 76 |
+
```python
|
| 77 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 78 |
+
model = AutoModelForCausalLM.from_pretrained("evalengine/unbound-e4b")
|
| 79 |
+
tok = AutoTokenizer.from_pretrained("evalengine/unbound-e4b")
|
| 80 |
+
```
|
| 81 |
|
| 82 |
## Acknowledgements
|
| 83 |
|
| 84 |
- Fine-tuned with [Unsloth](https://github.com/unslothai/unsloth) + Huggingface's [TRL](https://github.com/huggingface/trl).
|
| 85 |
- Abliteration via [heretic](https://github.com/p-e-w/heretic).
|
| 86 |
- Environment and training discipline ported from [autoresearch](https://github.com/karpathy/autoresearch).
|
| 87 |
+
|
| 88 |
+
## License
|
| 89 |
+
|
| 90 |
+
Apache-2.0, inherited from `google/gemma-4-E4B-it`.
|