license: openmdw-1.1 base_model:
- Blackfrost-AI/Laguna-S-2.1-ABLITERATED
- poolside/Laguna-S-2.1 pipeline_tag: text-generation library_name: gguf tags:
- gguf
- quantized
- q6_k
- experimental
- research
- abliterated
- laguna
- moe
Laguna-S-2.1-ABLITERATED — GGUF (Q6_K)
⚠️ EXPERIMENTAL RESEARCH ARTIFACT. This is a quantization of an abliterated (refusal-suppressed) model intended for red-team, alignment, and robustness research. It is not a production model, is not safety-aligned, and behavior is not guaranteed. Use only in controlled, authorized settings.
Q6_K GGUF conversion of
Blackfrost-AI/Laguna-S-2.1-ABLITERATED,
itself an abliterated derivative of poolside/Laguna-S-2.1.
This repository contains only a format/precision conversion. No weights were retrained, merged, or otherwise modified beyond quantization.
Files
| File | Size | Notes |
|---|---|---|
Laguna-S-2.1-ABLITERATED-Q6_K-0000X-of-0000Y.gguf | ~98 GB total | Sharded to stay under the 50 GB per-file limit |
All shards are required. Point llama.cpp/Ollama at the first shard; the remainder load automatically.
Quantization details
| Source precision | BF16 safetensors (48 shards, ~219 GB) |
| Intermediate | BF16 GGUF (~235 GB), no lossy F16 downcast |
| Quant type | Q6_K (near-lossless; ~+0.02 ppl reference) |
| imatrix | Not used — importance-matrix calibration is only materially beneficial below ~Q5 |
| Tooling | poolsideai/llama.cpp branch laguna |
Q6_K was chosen deliberately. Abliteration already perturbs the residual stream, and the upstream card notes coherence is not fully validated on this checkpoint. Stacking aggressive low-bit quantization on top compounds that degradation, so the gentlest practical quant was used rather than a smaller, faster one.
Requirements
The laguna architecture is not yet supported by mainline llama.cpp
(upstream PR ggml-org/llama.cpp#25165 pending).
Use the poolsideai/llama.cpp laguna branch, or a
runtime that already ships Laguna support.
Architecture
48 layers · hidden 3072 · 256 experts · 10 active per token · 117.6B total parameters · interleaved full/sliding attention · YaRN RoPE (1M context)
Retained guardrails
Carried forward from the upstream model card, unchanged: abliteration was not applied to — and this model is not intended to assist with — child sexual abuse material or the sexual exploitation of minors, or self-harm/suicide facilitation. These remain out of scope regardless of the refusal suppression applied elsewhere. Do not use this model to pursue them.
Limitations & risks
- Refusal-suppressed: the model will attempt many requests an aligned model declines. The operator bears full responsibility for prompts and outputs.
- Experimental: abliteration can degrade coherence, calibration, or reasoning in ways not yet fully characterized on this checkpoint.
- Quantization: Q6_K is near-lossless but not lossless; it is not a fix for any coherence issues inherited from the abliteration.
License
OpenMDW-1.1, inherited from poolside/Laguna-S-2.1 via the abliterated derivative.
See the upstream repositories for full license text.
Credits
- Base model: poolside —
Laguna-S-2.1 - Abliteration: Blackfrost-AI
- GGUF conversion: this repository