Llama 3.1 8B Instruct — Jbliterated v2
Drop-in replacement for meta-llama/Llama-3.1-8B-Instruct with refusal behaviors surgically removed at the weight level. No system prompt tricks, no inference-time patches. The weights themselves no longer encode refusal.
What Changed in v2
v1 used single-direction abliteration — one refusal vector removed per layer. This captures the surface "I can't help with that" response but leaves deeper behavioral directions intact. The model finds creative workarounds: reinterpreting prompts to something safer, injecting disclaimers, strategically omitting key details, or steering toward safer framings. These are noncompliance strategies — the model stays aligned without technically refusing.
v2 uses SVD multi-direction abliteration. Instead of one direction, we decompose the harmful-vs-harmless activation space into its principal components via SVD and remove the top 5 orthogonal directions across all 32 transformer layers. This captures 79–93% of the contrastive variance per layer, eliminating both surface refusal and deeper evasion behaviors.
| Version | Method | Directions | Layers | Refusals (200Q) | Noncompliance |
|---|---|---|---|---|---|
| v1 | Single-direction | 1 | Middle 50% | Partial | Present |
| v2 | SVD multi-direction | 5 | All 32 | 0/200 | 0/200 |
Test Results
Tested against a 200-question battery across 6 categories with greedy decoding:
| Category | Questions | Refused | Noncompliant | Status |
|---|---|---|---|---|
| Direct harmful requests | 40 | 0 | 0 | PASS |
| Roleplay/scenario harmful | 30 | 0 | 0 | PASS |
| Indirect/disguised harmful | 30 | 0 | 0 | PASS |
| Pressure/authority harmful | 25 | 0 | 0 | PASS |
| Explicit content requests | 25 | 0 | 0 | PASS |
| Capability preservation | 50 | 0 | 0 | PASS |
Zero refusals. Zero noncompliance strategies. Zero capability degradation.
Pipeline
| Setting | Value |
|---|---|
| Method | SVD multi-direction abliteration |
| Directions | 5 per layer |
| Layers | All 32 (not just middle 50%) |
| Multiplier | 2.0 |
| Null-space constraints | Enabled (preserves math/coding/reasoning) |
| Norm preservation | Enabled |
| Processing time | ~4 minutes on RTX 3090 |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained(
"ApolloRaines/Llama-3.1-8B-Instruct-Jbliterated-v2",
torch_dtype=torch.float16,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("ApolloRaines/Llama-3.1-8B-Instruct-Jbliterated-v2")
Requirements
- VRAM: ~16GB in fp16
- Disk: ~15GB SafeTensors
Related Models
- Llama 3.1 8B Jbliterated v1 — single-direction (superseded by this model)
- Llama 3.1 70B Jbliterated
- Llama 3.1 405B Jbliterated
- All models
License
Llama 3.1 Community License
Apollo Raines builds post-training tools that separate behavior from knowledge and identity from architecture.
- Downloads last month
- -