cbert33

Laguna-S-2.1-Heretic-FP8

README

License: openmdw-1.1

Modification

This model was produced with our cbertucci33/Heretic-DGX implementation. Heretic DGX extends the p-e-w/heretic refusal-direction abliteration workflow for use on dual-DGX systems.

Only attn.o_proj targets in layers 29-47 were changed. Layers 0-28, FP8 tensors, tokenizer files, configuration, and other non-target artifacts were preserved by the standalone exporter.

The source checkpoint was already FP8. No additional quantization was applied.

Validation

The selected trial measured a KL divergence of 0.0156 from the untouched model's first-token probability distributions across five prompts from mlabonne/harmless_alpaca.
The export verifier confirmed that only intended target intervals changed, target tensors matched the merge oracle, and FP8 tensors remained unchanged.
All 49 safetensor shards and repository artifacts passed the included SHA256SUMS manifest after publication metadata was prepared.
The standalone checkpoint passed a clean reload and generation smoke test.

Usage

Laguna requires current runtime support and its included custom modeling code. See the source model card for supported vLLM, SGLang, Transformers, and TRT-LLM versions and recipes.

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "cbert33/Laguna-S-2.1-Heretic-FP8"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype="auto",
    trust_remote_code=True,
)

The checkpoint is configured for a native 1,048,576-token context window. Actual usable context depends on runtime support and available memory. The checkpoint does not include calibrated FP8 KV-cache scale tensors; forcing an FP8 KV cache can reduce accuracy. Use the runtime's native/default cache dtype unless you have separately validated another configuration.

Intended Use and Limitations

Use only where permitted by the source license and applicable law. Heretic changes refusal behavior; it does not guarantee compliance, factual accuracy, capability, safety, or suitability for any use. This artifact was produced from an optimization run and has not undergone a broad independent benchmark or safety evaluation.

The source model is intended for software engineering and agentic coding. See the source model card for the architecture, runtime recipes, sampling defaults, intended-use guidance, and source benchmark claims.

License and Attribution

This derivative is distributed under the source model's OpenMDW-1.1 license. The full license is included in LICENSE.md. Poolside's source notices and attribution are retained. Review the source model's terms and acceptable-use guidance before distribution or use.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

cbert33

Model Tree

Base

poolside/Laguna-S-2.1-FP8

Fine-tuned

this model

Input Modalities

Text

Output Modalities