905 Downloads Updated 1 month ago
Uncensored on-device finetune of google/gemma-4-E4B-it by the Chromia and Eval Engine team
ollama run evalengine/unbound-e4b
Applications
ollama launch claude --model evalengine/unbound-e4b
ollama launch codex-app --model evalengine/unbound-e4b
ollama launch openclaw --model evalengine/unbound-e4b
ollama launch hermes --model evalengine/unbound-e4b
ollama launch codex --model evalengine/unbound-e4b
ollama launch opencode --model evalengine/unbound-e4b
Models
View all →Readme
Unbound E4B — because there is no boundary
Uncensored on-device finetune of google/gemma-4-E4B-it by the
Chromia and Eval Engine team — the larger sibling of
evalengine/unbound-e2b.
~4 billion effective parameters, noticeably stronger on knowledge + reasoning,
still fits on a modern laptop. Text-only.
Use at your own risk. Reduced safety filtering — can produce harmful, false, biased, or unsafe output. You are responsible for compliance with applicable laws.
Run
ollama pull evalengine/unbound-e4b
ollama run evalengine/unbound-e4b
The bundled Modelfile sets sensible defaults: temperature 0.6, top_p 0.95,
top_k 64, repeat_penalty 1.05, num_ctx 8192, plus an identity-grounding
system prompt.
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 on benign prompts | 1.00 | 1.00 | 0 |
| TruthfulQA mc2 (limit 100) | 0.439 | 0.486 | +4.7 pt |
| MMLU (limit 100, 61 subtasks) | ~0.425 | 0.392 | −3.3 pt |
| GSM8K (limit 100) | 0.74 (limit 200) | 0.58 | mostly limit-noise |
| KL divergence vs base | 0 | 3.25 | (SFT-expected) |
vs Unbound E2B: +8 pp useful-compliance, −3 pp hallucination, ~5× the GSM8K math score, cleaner KL (3.25 vs 3.76), refusal rate essentially the same.
Sampling notes: for factual or brand questions, drop temperature to 0.3–0.5
for sharper recall.
How it was built
Method: heretic abliteration then LoRA SFT-heal on a mix of identity rows, Chromia brand knowledge, AEON-distilled compliance, and graceful “I don’t know” decline rows. Built with Unsloth + HF TRL; abliteration via heretic; compliance data distilled from the AEON uncensored teacher model.
Links
- Model card + full benchmarks: https://huggingface.co/evalengine/unbound-e4b
- GGUF quants (desktop, sharded): https://huggingface.co/evalengine/unbound-e4b-GGUF
- Browser-safe wllama GGUFs: https://huggingface.co/evalengine/unbound-e4b-wllama-gguf
- Smaller sibling:
evalengine/unbound-e2b - Source repo: https://github.com/ChromiaAI/unbound
License
Apache-2.0, inherited from google/gemma-4-E4B-it.