Llama-3.1-8B-Lexi-Uncensored-V2

Upstream

Llama-3.1-8B-Instruct

Modified by

Orenguteng

Method

finetune

Canonical modified weights

available upstream (high-precision weights published)

Method (as documented): finetune Method parameters: undocumented — no dataset, no training recipe, no LoRA-vs-full, no hyperparameters disclosed Modifier: Orenguteng Notes and caveats: - modification_kind_note: CONFIRMED, not abliteration. The card describes a fine-tuned adaptation and explicitly distinguishes it from abliteration; the repo carries no 'abliterated' tag and no abliteration vocabulary. The source document's suspicion is upheld. - gguf_warning: Card warns Q4 has refusal regressions and recommends F16 or Q8. This is a direct preservation argument for keeping the high-precision weights rather than relying on small quants. - lineage_provenance_reason: Modification kind is clearly stated, but with no base_model metadata, no dataset and no recipe, nothing about the weights' provenance is machine-verifiable. Conflicting claims (each stored separately, deliberately NOT reconciled): [1] Author claims V2 is 'more compliant' and 'smarter' than V1 — no numbers. [2] Author separately claims Q4 quantization REINTRODUCES refusals — informal, unquantified, and in tension with the general assumption that quants preserve behavior. Record both; do not reconcile. Source: refusal-modified-acquisition-manifest.json v1.0.0 (research snapshot 2026-09-20), candidate 84.

Explore further