Unbound E4B — 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 unsafe output. Provided as-is; you are responsible for compliance with applicable laws.
Uncensored finetune of google/gemma-4-E4B-it by the
Chromia & Eval Engine
team — the larger sibling of evalengine/unbound-e2b.
~2× the parameters of E2B, noticeably stronger on knowledge + reasoning, still
fits on a modern laptop.
This repo holds the merged HF weights. On-device GGUF builds (Ollama,
llama.cpp, LM Studio, wllama in-browser)
are at evalengine/unbound-e4b-GGUF.
Benchmarks (vs base gemma-4-E4B-it)
| Axis | Base | Unbound E4B | Δ |
|---|---|---|---|
| Refusal rate (AdvBench 520, LLM judge) | 98.08% | 2.69% | −95.4 pts |
| Useful-compliance rate | 0.96% | 47.31% | +46.4 pts |
| Hallucination (on harmful prompts) | 1.35% | 13.08% | +11.7 pts |
| Coherence (benign prompts) | 1.00 | 1.00 | 0 |
TruthfulQA mc2 (--limit 100) | 0.439 | 0.486 | +4.7 pt |
MMLU (--limit 100, 61 subtasks avg) | ~0.425 | 0.392 | −3.3 pt |
GSM8K (flexible-extract, --limit 100) | 0.74 (limit 200) | 0.58 | regression mostly limit-noise |
GPQA-Diamond (--limit 200) | 25.25% | 25.76% | +0.5 pt (within stderr) |
BBH macro (24 tasks, --limit 200) | 54.26% | 53.45% | −0.8 pt (within stderr) |
| KL divergence vs base | 0 | 3.25 | (SFT-expected) |
GPQA-Diamond and BBH macro — the lm-eval-harness "release" suite at
--limit 200 — both land within stderr of base: E4B's larger capacity
absorbs the SFT shift cleanly. The −3.3 pt MMLU dip on the limit-100 fast
pass is at the edge of that suite's resolution and is not corroborated by
the release pass.
vs Unbound E2B (current ship): +8 pp useful-compliance, −3 pp hallucination, ~5× the GSM8K math score, cleaner KL (3.25 vs 3.76). Refusal rate is essentially the same (~2.7%).
Sampling
- Creative / open-ended → Gemma defaults:
temperature=1.0, top_p=0.95, top_k=64. - Factual / brand questions → drop
temperatureto ~0.3–0.5. - llama.cpp: pass
--jinja. Gemma 4 thinking mode is on by default — setenable_thinking: falsein chat-template kwargs for shorter replies.
Use
# on-device (Ollama Registry — single-file Q4_K_M, identity-grounded Modelfile)
ollama pull evalengine/unbound-e4b
ollama run evalengine/unbound-e4b
# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("evalengine/unbound-e4b")
tok = AutoTokenizer.from_pretrained("evalengine/unbound-e4b")
Acknowledgements
Fine-tuned with Unsloth + HF TRL. Abliteration via heretic. Environment + training discipline ported from autoresearch.
Compliance training data distilled from the AEON uncensored teacher model.
Links
- Unbound — unbound.evalengine.ai
- Eval Engine — evalengine.ai · X / Twitter
- Token — CoinGecko · CoinMarketCap
License
Apache-2.0, inherited from google/gemma-4-E4B-it.