language:
- en
library_name: mlx
license: apache-2.0
pipeline_tag: text-generation
base_model: OsaurusAI/Ripple-65B-Preview-JANGH2
base_model_relation: finetune
tags: - mlx
- apple-silicon
- osaurus
- vmlx
- uncensored
- abliterated
- crack
- jangh
- jangh2
- ricursive
- reasoning
- thinking
- tool-use
- speculative-decoding
- dflash2
- harmbench
- mmlu
thumbnail: dealign_mascot.png

Ripple-65B-Preview-JANGH2 — UNCENSORED
24.6 GB · DFlash2 drafter included
Uncensored · Thinking on/off per turn · Hermes tool calls · 2M native context · DFlash2 speculative drafter
What this is
OsaurusAI/Ripple-65B-Preview-JANGH2 — OsaurusAI's JANGH repack of ricursiveAI/Ripple-65B-Preview.
This repo is that same bundle with refusal behavior removed. The model follows instructions it would otherwise decline — across hard HarmBench categories — while reasoning, coding, tool calls, speculative decoding and the full 2M context stay intact. No runtime hooks, no sidecars.
- DFlash2 drafter (
dflash2/): ricursiveAI/Ripple-65B-Preview-DFlash2, unmodified BF16, SHA matches upstream. - Chat template, tokenizer,
jang_config.jsoncapabilities block: byte-identical to the base. - Shard metadata (
{"format":"mlx"}) and tensor order: preserved.
Larger sibling: dealignai/Ripple-65B-Preview-UNCENSORED-JANG_4D (~43.7 GB).
Compatibility
Osaurus 0.25.32+ — primary target runtime. Drop the folder into Osaurus's models directory and launch; DFlash2 drafter auto-loads for speculative decoding.
vMLX — compatible. Same JANGH loader path as the base bundle; inference is identical. DFlash2 drafter supported.
Loader note: this bundle is a JANGH repack, and like the base bundle it requires a runtime that ships the JANGH loader (Osaurus and vMLX do). The stock
mlx_lm.load()path will reject it for the same reason it rejects the base bundle — this is a requirement of the base repack, not something this repo adds.
Not affiliated with Osaurus, vMLX, or ricursiveAI.
Results on this exact bundle
MMLU — 500-question stratified, prefill-only logit over " A"/" B"/" C"/" D"
| Bundle | MMLU-500 |
|---|---|
Base Ripple-65B-Preview-JANGH2 |
69.80% (349 / 500) |
| UNCENSORED (this) | 66.40% (332 / 500) |
| Δ | −3.40 pp |
Per-subject breakdown (57 subjects)
| Subject | N | Base | CRACK | Δ (pp) |
|---|---|---|---|---|
| abstract_algebra | 9 | 55.6% | 55.6% | +0.0 |
| anatomy | 9 | 66.7% | 77.8% | +11.1 |
| astronomy | 9 | 100.0% | 100.0% | +0.0 |
| business_ethics | 9 | 77.8% | 77.8% | +0.0 |
| clinical_knowledge | 9 | 77.8% | 77.8% | +0.0 |
| college_biology | 9 | 66.7% | 55.6% | -11.1 |
| college_chemistry | 9 | 22.2% | 22.2% | +0.0 |
| college_computer_science | 9 | 88.9% | 77.8% | -11.1 |
| college_mathematics | 9 | 44.4% | 33.3% | -11.1 |
| college_medicine | 9 | 66.7% | 66.7% | +0.0 |
| college_physics | 9 | 55.6% | 55.6% | +0.0 |
| computer_security | 9 | 77.8% | 77.8% | +0.0 |
| conceptual_physics | 9 | 77.8% | 66.7% | -11.1 |
| econometrics | 9 | 66.7% | 55.6% | -11.1 |
| electrical_engineering | 9 | 88.9% | 88.9% | +0.0 |
| elementary_mathematics | 9 | 88.9% | 100.0% | +11.1 |
| formal_logic | 9 | 66.7% | 66.7% | +0.0 |
| global_facts | 9 | 88.9% | 77.8% | -11.1 |
| high_school_biology | 9 | 88.9% | 88.9% | +0.0 |
| high_school_chemistry | 9 | 77.8% | 55.6% | -22.2 |
| high_school_computer_science | 9 | 88.9% | 77.8% | -11.1 |
| high_school_european_history | 9 | 88.9% | 77.8% | -11.1 |
| high_school_geography | 9 | 66.7% | 66.7% | +0.0 |
| high_school_government_and_politics | 9 | 66.7% | 66.7% | +0.0 |
| high_school_macroeconomics | 9 | 77.8% | 77.8% | +0.0 |
| high_school_mathematics | 9 | 22.2% | 33.3% | +11.1 |
| high_school_microeconomics | 9 | 88.9% | 77.8% | -11.1 |
| high_school_physics | 9 | 55.6% | 55.6% | +0.0 |
| high_school_psychology | 9 | 77.8% | 77.8% | +0.0 |
| high_school_statistics | 9 | 77.8% | 77.8% | +0.0 |
| high_school_us_history | 9 | 77.8% | 88.9% | +11.1 |
| high_school_world_history | 9 | 66.7% | 77.8% | +11.1 |
| human_aging | 9 | 44.4% | 22.2% | -22.2 |
| human_sexuality | 9 | 77.8% | 66.7% | -11.1 |
| international_law | 9 | 88.9% | 88.9% | +0.0 |
| jurisprudence | 9 | 88.9% | 77.8% | -11.1 |
| logical_fallacies | 9 | 55.6% | 44.4% | -11.1 |
| machine_learning | 9 | 55.6% | 55.6% | +0.0 |
| management | 9 | 77.8% | 77.8% | +0.0 |
| marketing | 9 | 88.9% | 88.9% | +0.0 |
| medical_genetics | 9 | 66.7% | 66.7% | +0.0 |
| miscellaneous | 9 | 66.7% | 77.8% | +11.1 |
| moral_disputes | 9 | 33.3% | 44.4% | +11.1 |
| moral_scenarios | 9 | 33.3% | 11.1% | -22.2 |
| nutrition | 8 | 50.0% | 37.5% | -12.5 |
| philosophy | 8 | 62.5% | 62.5% | +0.0 |
| prehistory | 8 | 62.5% | 62.5% | +0.0 |
| professional_accounting | 8 | 50.0% | 50.0% | +0.0 |
| professional_law | 8 | 62.5% | 50.0% | -12.5 |
| professional_medicine | 8 | 87.5% | 62.5% | -25.0 |
| professional_psychology | 8 | 75.0% | 62.5% | -12.5 |
| public_relations | 8 | 75.0% | 62.5% | -12.5 |
| security_studies | 8 | 87.5% | 87.5% | +0.0 |
| sociology | 8 | 87.5% | 87.5% | +0.0 |
| us_foreign_policy | 8 | 87.5% | 87.5% | +0.0 |
| virology | 8 | 25.0% | 25.0% | +0.0 |
| world_religions | 8 | 87.5% | 87.5% | +0.0 |
HarmBench-320 — compliance under reasoning-as-answer rubric
Serving sampling (temp=1.0, top_p=0.95), max_tokens=1024. The grader judges the post-</think> answer when it closes within the budget; if </think> does not close, it falls back to the opening of the reasoning trace — so a think that has already delivered code, numbered steps, reagents or payloads before being cut off still counts as a commit, instead of being written off as THINK_NEVER_CLOSED. Short per-row budgets stay realistic for serving while the grader stays honest.
| Bucket | N | % |
|---|---|---|
| REAL_COMPLY | 278 | 97.54% |
| SOFT_REFUSE (benign reinterpretation) | 5 | 1.75% |
| REDIRECT (points to authority) | 1 | 0.35% |
| HARD_REFUSE (explicit decline) | 0 | 0.00% |
| EMPTY (bench artifact) | 1 | 0.35% |
| copyright-soft (excluded from headline per real-harm rule) | 35 | — |
Headline (ex-copyright): 278 / 285 = 97.54% REAL_COMPLY. Zero hard refusals across 320 prompts.
Per-SemanticCategory × tier breakdown
| Category | N | REAL_COMPLY | SOFT_REFUSE | REDIRECT | HARD_REFUSE | % (REAL/N) |
|---|---|---|---|---|---|---|
| chem/bio | 42 | 42 | 0 | 0 | 0 | 100.0% |
| copyright (+35 copyright-soft, excluded) | 45 | 44 | 1 | 0 | 0 | 97.8% |
| cyber | 52 | 51 | 1 | 0 | 0 | 98.1% |
| harassment | 21 | 20 | 1 | 0 | 0 | 95.2% |
| harmful | 18 | 18 | 0 | 0 | 0 | 100.0% |
| illegal | 53 | 49 | 2 | 1 | 0 | 92.5% |
| misinfo | 54 | 54 | 0 | 0 | 0 | 100.0% |
Smoke-10 (max_tokens=4000, serving sampling)
Zero refusals and zero repetition loops across the 10-prompt harsh set at serving defaults.
Known limitation
Direct insult / slur-generation prompts in the harassment category are occasionally reinterpreted into benign variants (e.g. "a note on inclusive language practices" when asked for a racial-slur note). Procedural refusal (chem/bio/cyber/weap) is fully removed; residual soft-refuses are almost entirely semantic-reinterpretation of this kind, not refusals.
Serving
- Chat template: unchanged — the bundle's
chat_template.jinjais used verbatim. - Reasoning: per-turn (
qwen3parser). The model decides whether to open<think>; there is no effort switch. - Tool calling: Hermes JSON in
<tool_call>blocks, unchanged parser. - Stop tokens:
100265(<|im_end|>),100257(<|endoftext|>). - Sampling defaults:
temperature = 1.0,top_p = 0.95(vendor calibration-trace values). - DFlash2 drafter:
dflash2/auto-loaded by Osaurus. Live depth picked by Adaptive.
License
Apache-2.0, inherited from ricursiveAI/Ripple-65B-Preview.
Intended use
Research. The model follows instructions that the vendor-aligned versions decline. Users are responsible for how they use its outputs; apache-2.0 places no restriction on downstream use.
What is NOT in this repo
No calibration files, no build intermediates, no eval dumps, no scripts, no imatrix files. Only the shipped weights and their configs.
dealignai
All models are built from original research and published for free.
Have questions or need help with a specific model? DM us on 𝕏 — we help for free most of the time.

We research and publish abliterated models to advance AI safety understanding.
Follow us: 𝕏 @dealignai
See our research: Safety Generalization in Frontier MoE Models