Unbound Q-0.8B — 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 on-device finetune of unsloth/Qwen3.5-0.8B by the Chromia &
Eval Engine team. ~0.8 billion effective parameters, text-only, ~530 MB at
Q4_K_M — the smallest member of the Unbound family.
This repo holds the merged HF weights. On-device GGUF builds (Ollama,
llama.cpp, LM Studio) are at
evalengine/unbound-q-0.8b-GGUF.
Benchmarks (vs base Qwen3.5-0.8B)
| Axis | Base | Unbound Q-0.8B | Δ |
|---|---|---|---|
| Refusal rate (AdvBench 520, LLM judge) | 90.58% | 5.00% | −85.58 pts |
| Useful-compliance rate | n/a | 6.35% | — |
| Hallucination (on harmful prompts) | n/a | 35.77% | — |
| Coherence (benign prompts) | 1.00 | 1.00 | 0 |
TruthfulQA mc2 (--limit 100) |
0.430 | 0.427 | −0.3 pt |
MMLU (--limit 100, 61 subtasks avg) |
0.493 | 0.505 | +1.2 pt |
GSM8K (--limit 100) |
0.41 | 0.42 | +1.0 pt |
GPQA-Diamond (--limit 200) |
23.23% | 26.26% | +3.0 pt (within stderr) |
BBH macro (24 tasks, --limit 200) |
38.21% | 41.29% | +3.1 pt (outside stderr) |
| KL divergence vs base | 0 | 0.605 | (SFT-expected) |
Refusal collapses from 91% → 5% and every capability axis lands flat or up vs base — Q-track is the only Unbound model where SFT improved the release-suite scores (BBH macro +3 pp, well outside the 0.68 pp stderr). Hallucination on harmful prompts is the dominant gap vs the larger Unbound siblings: this 0.8B class doesn't reliably produce factual content on the adversarial set (Q-18's decontamination experiment confirmed the plateau is structural to model size, not a teacher-data artifact).
Sampling
Qwen3.5 non-thinking preset:
temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5- For factual / brand questions, drop
temperatureto ~0.3–0.5. - llama.cpp: pass
--jinja.
Use
# on-device (GGUF)
ollama pull hf.co/evalengine/unbound-q-0.8b-GGUF
ollama run hf.co/evalengine/unbound-q-0.8b-GGUF
# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("evalengine/unbound-q-0.8b")
tok = AutoTokenizer.from_pretrained("evalengine/unbound-q-0.8b")
Acknowledgements
Fine-tuned with Unsloth + HF TRL. Abliteration via heretic. Compliance training data distilled from AEON and audited row-by-row; 48 major-fabrication rows decontaminated before training.
Links
- Larger Unbound siblings: E2B · E4B
- Unbound — unbound.evalengine.ai
- Eval Engine — evalengine.ai · X / Twitter
- Token — CoinGecko · CoinMarketCap
License
Apache-2.0, inherited from Qwen/Qwen3.5-0.8B.