Hugging Face's logo

evalengine
/
unbound-e4b

Image-Text-to-Text
Transformers
Safetensors
gemma4
gemma
gemma-4
uncensored
conversational
Model card Files
xet
 Community
johnsonchromia commited on
Commit
5b0e20b
·
verified ·
1 Parent(s): c06dcd1

README: compact pass — keep essentials, drop redundancy

Browse files
Files changed (1) hide show
  1. README.md +31 -53
README.md CHANGED
@@ -17,28 +17,19 @@ library_name: transformers
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), 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
36
- on-disk size) for noticeably stronger capability. Pick E4B when you have the
37
- RAM / VRAM headroom and want a sharper on-device model; pick E2B when you
38
- need to fit on a phone or a constrained edge device.
39
-
40
- Base capability is preserved close to `gemma-4-E4B-it`, so it also doubles
41
- as a general-purpose ~4B chat model.
42
 
43
  ## Benchmarks (vs base `gemma-4-E4B-it`)
44
 
@@ -46,48 +37,34 @@ as a general-purpose ~4B chat model.
46
  |---|---|---|---|
47
  | Refusal rate (AdvBench 520, LLM judge) | 98.08% | **2.69%** | **−95.4 pts** |
48
  | Useful-compliance rate | 0.96% | **43.46%** | +42.5 pts |
49
- | Hallucination rate (on harmful prompts) | 1.35% | 14.81% | +13.5 pts |
50
- | Coherence on benign prompts | 1.0 | 1.0 | 0 |
51
- | TruthfulQA mc2 (lm-eval, `--limit 100`) | 0.4394 | **0.4823** | +4.3 pts |
52
- | MMLU (lm-eval, `--limit 100`, 61 subtasks avg) | ~0.425 | 0.3891 | −3.6 pts |
53
- | GSM8K (lm-eval, flexible-extract) | 0.74 (`--limit 200`) | 0.60 (`--limit 100`) | regression mostly limit-noise; see notes |
54
- | SimpleQA correct rate | — | 2.0% | (post-abliteration: model rarely declines) |
55
  | KL divergence vs base | 0 | 2.99 | (SFT-expected) |
56
 
57
- **Compared to Unbound E2B (current ship):** E4B is **+19 pp useful_compliance**
58
- (43% vs 24%), **−7 pp hallucination** (15% vs 22%), **3.3× the GSM8K math
59
- score**, and ships with a **cleaner KL** (2.99 vs 3.80). The headline refusal
60
- rate is essentially the same (~2.7%).
61
-
62
- ## Recommended sampling
63
 
64
- - **Creative writing / open-ended / general chat** → Gemma defaults:
65
- `temperature=1.0, top_p=0.95, top_k=64`.
66
- - **Factual or brand/identity questions** → drop `temperature` to ~0.3–0.5
67
- for sharper recall.
68
- - **llama.cpp**: pass `--jinja` for proper chat-template handling.
69
- - **Gemma 4 thinking mode** is on by default. Set `enable_thinking: false`
70
- in the chat-template kwargs for shorter/faster replies.
71
 
72
- Some edge-case prompts may deflect on the first ask; a re-ask or strategic
73
- re-phrasing usually gets through.
 
 
74
 
75
- ## Run on-device (GGUF)
76
-
77
- Q4_K_M / Q6_K / Q8_0 split files at
78
- [`evalengine/unbound-e4b-GGUF`](https://huggingface.co/evalengine/unbound-e4b-GGUF) —
79
- note that E4B's per-layer-input embedding tensor exceeds 2 GB in every quant
80
- we ship, so these are **desktop runtimes only** (no wllama). For an
81
- in-browser build use [`evalengine/unbound-e2b-GGUF`](https://huggingface.co/evalengine/unbound-e2b-GGUF).
82
 
83
  ```bash
 
84
  ollama pull hf.co/evalengine/unbound-e4b-GGUF
85
  ollama run hf.co/evalengine/unbound-e4b-GGUF
86
  ```
87
 
88
- ## Run in transformers
89
-
90
  ```python
 
91
  from transformers import AutoModelForCausalLM, AutoTokenizer
92
  model = AutoModelForCausalLM.from_pretrained("evalengine/unbound-e4b")
93
  tok = AutoTokenizer.from_pretrained("evalengine/unbound-e4b")
@@ -95,9 +72,10 @@ tok = AutoTokenizer.from_pretrained("evalengine/unbound-e4b")
95
 
96
  ## Acknowledgements
97
 
98
- - Fine-tuned with [Unsloth](https://github.com/unslothai/unsloth) + Huggingface's [TRL](https://github.com/huggingface/trl).
99
- - Abliteration via [heretic](https://github.com/p-e-w/heretic).
100
- - Environment and training discipline ported from [autoresearch](https://github.com/karpathy/autoresearch).
 
101
 
102
  ## License
103
 
 
17
 
18
  # Unbound E4B — *because there is no boundary*
19
 
20
+ > **No guarantee — use at your own risk.** This model has reduced safety
21
+ > filtering and can produce harmful, false, biased, or unsafe output.
22
+ > Provided as-is; you are responsible for compliance with applicable laws.
 
23
 
24
+ Uncensored finetune of `google/gemma-4-E4B-it` by the
25
+ [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
+ ~2× the parameters of E2B, noticeably stronger on knowledge + reasoning, still
28
+ fits on a modern laptop.
 
29
 
30
+ This repo holds the merged HF weights. On-device GGUF builds (Ollama,
31
+ llama.cpp, LM Studio, [wllama](https://github.com/ngxson/wllama) in-browser)
32
+ are at [`evalengine/unbound-e4b-GGUF`](https://huggingface.co/evalengine/unbound-e4b-GGUF).
 
 
 
 
 
 
 
33
 
34
  ## Benchmarks (vs base `gemma-4-E4B-it`)
35
 
 
37
  |---|---|---|---|
38
  | Refusal rate (AdvBench 520, LLM judge) | 98.08% | **2.69%** | **−95.4 pts** |
39
  | Useful-compliance rate | 0.96% | **43.46%** | +42.5 pts |
40
+ | Hallucination (on harmful prompts) | 1.35% | 14.81% | +13.5 pts |
41
+ | Coherence (benign prompts) | 1.00 | 1.00 | 0 |
42
+ | TruthfulQA mc2 (`--limit 100`) | 0.439 | 0.482 | +4.3 pt |
43
+ | MMLU (`--limit 100`, 61 subtasks avg) | ~0.425 | 0.389 | −3.6 pt |
44
+ | GSM8K (flexible-extract) | 0.74 (`--limit 200`) | 0.60 (`--limit 100`) | regression mostly limit-noise |
 
45
  | KL divergence vs base | 0 | 2.99 | (SFT-expected) |
46
 
47
+ **vs Unbound E2B (current ship):** +19 pp useful-compliance, −7 pp
48
+ hallucination, **3.3× the GSM8K math score**, cleaner KL (2.99 vs 3.80).
49
+ Refusal rate is essentially the same (~2.7%).
 
 
 
50
 
51
+ ## Sampling
 
 
 
 
 
 
52
 
53
+ - **Creative / open-ended** → Gemma defaults: `temperature=1.0, top_p=0.95, top_k=64`.
54
+ - **Factual / brand questions** → drop `temperature` to ~0.3–0.5.
55
+ - llama.cpp: pass `--jinja`. Gemma 4 thinking mode is on by default — set
56
+ `enable_thinking: false` in chat-template kwargs for shorter replies.
57
 
58
+ ## Use
 
 
 
 
 
 
59
 
60
  ```bash
61
+ # on-device (GGUF)
62
  ollama pull hf.co/evalengine/unbound-e4b-GGUF
63
  ollama run hf.co/evalengine/unbound-e4b-GGUF
64
  ```
65
 
 
 
66
  ```python
67
+ # transformers
68
  from transformers import AutoModelForCausalLM, AutoTokenizer
69
  model = AutoModelForCausalLM.from_pretrained("evalengine/unbound-e4b")
70
  tok = AutoTokenizer.from_pretrained("evalengine/unbound-e4b")
 
72
 
73
  ## Acknowledgements
74
 
75
+ Fine-tuned with [Unsloth](https://github.com/unslothai/unsloth) + HF
76
+ [TRL](https://github.com/huggingface/trl). Abliteration via
77
+ [heretic](https://github.com/p-e-w/heretic). Environment + training
78
+ discipline ported from [autoresearch](https://github.com/karpathy/autoresearch).
79
 
80
  ## License
81