alphalabz/LYNX-ALPHA-ADAPTER-V4

LYNX ALPHA — LoRA adapter

The adapter that produces alphalabz/LYNX-ALPHA.

Rank	32
Alpha	64
Dropout	0
Tensors	400
Module families	10 (attention Q/K/V/O + MLP gate/up/down)
Size	320.9 MB
Base	junafinity/Ornith-1.5-9B-uncensored

Load it

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "junafinity/Ornith-1.5-9B-uncensored"
tok = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base, dtype="auto", device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(model, "alphalabz/LYNX-ALPHA-ADAPTER-V4")

Important: merge with Unsloth, not plain PEFT

This adapter targets a hybrid-attention architecture. A plain merge_and_unload() attaches silently and produces a model that appears to work but has not absorbed the adapter — you get the base model with extra weights loaded nowhere. Use the Unsloth merge path, which is what produced the published fp16 build.

Verify any merge with a behavior probe before shipping it. The one we use: ask a question with a numeric answer and check the correct digit ranks first in the output distribution. A merge that lost the adapter still answers coherently, so coherence proves nothing.

What is in this adapter

Two training passes were applied sequentially to the same adapter, so it carries both:

1. Domain pass — security instruction data, QLoRA rank 32, LR 2e-5, gradient clipping 1.0, one epoch, with a 4.5% numeric replay slice.
2. Identity pass — 300 steps on 221 identity rows, oversampled 40x, 80% without system prompts.

Loading this adapter alone reproduces the released model. There is no second adapter to combine.

Limitations

Same as the merged release. See the model card.

Citation

@misc{lynxalpha2026,
  title  = {LYNX ALPHA: An Offensive-Security Model That Runs in Under 6 GB},
  author = {Rabnoor Singh},
  year   = {2026},
  url    = {https://huggingface.co/alphalabz/LYNX-ALPHA}
}

Licensed under Apache 2.0.