Llama-3.1-70B-Instruct-lorablated

Upstream

Llama-3.1-70B-Instruct

Modified by

mlabonne

Method

lorablated

Canonical modified weights

available upstream (high-precision weights published)

Method (as documented): lorablated Method parameters: LoRA rank 64; merge_method task_arithmetic; dtype bfloat16. mergekit_config.yml is in the repo and is the full spec — preserve it. Modifier: mlabonne Parent lineage: Weights have one link, but the MODIFICATION VECTOR carries cross-generation provenance: a failspy-style Llama-3 (not 3.1) 70B abliteration -> rank-64 LoRA extraction -> applied to Llama-3.1-70B. The schema cannot express this second axis. Notes and caveats: - base_model_import_note: imported during this run - modification_kind_note: DUAL CLASSIFICATION — mergekit task-arithmetic is the literal operation; the intent and effect are abliteration. Repo carries both 'abliterated' and 'mergekit' tags. Recorded as merge (mechanism); if the catalog prefers effect, abliteration is defensible. - gguf_note: Card points to a bartowski GGUF with IQ quants as better-featured; exact bartowski repo ID unresolved. - evaluation_note: NONE. Card claims suitability for general use and roleplay without supporting figures. Source: refusal-modified-acquisition-manifest.json v1.0.0 (research snapshot 2026-09-20), candidate 82.

Explore further