Manitec/Plex_Nyhex_v1

Plex Nyhex v1

Plex Nyhex v1 is a LoRA/PEFT adapter for Manitec/Qwen3-4B-Instruct-2507-uncensored-v2. This repository also includes a GGUF artifact for use with compatible local inference runtimes.

The adapter and GGUF are separate loading paths. The adapter requires the base model; it is not a complete standalone Transformers checkpoint. The GGUF's quantization, conversion history, and inclusion of this adapter have not been documented or independently verified.

Repository contents

File	Purpose
adapter_model.safetensors	LoRA adapter weights; approximately 132 MB.
adapter_config.json	Base-model reference and PEFT configuration.
Plex_Nyhex_q4_v1.gguf	GGUF artifact; approximately 2.50 GB, decimal units.
tokenizer.json	Exported tokenizer data.
tokenizer_config.json	Tokenizer configuration.
chat_template.jinja	Exported chat-formatting template.
README.md	Repository documentation and Hugging Face model card.
.gitattributes	Git attributes and large-file tracking rules.

Adapter configuration

Values below are read from adapter_config.json; they are not a complete training record.

Setting	Value
Base model	Manitec/Qwen3-4B-Instruct-2507-uncensored-v2
Architecture reference	Qwen3ForCausalLM
Adapter type	LORA
Task type	CAUSAL_LM
Rank (r)	16
LoRA alpha	32
LoRA dropout	0
Bias	none
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Recorded PEFT version	0.19.1
Inference mode	true

Load the adapter

Install compatible versions of PyTorch, Transformers, PEFT, and Accelerate:

pip install torch transformers peft accelerate

Example loading path (not yet validated against this repository):

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE_MODEL = "Manitec/Qwen3-4B-Instruct-2507-uncensored-v2"
ADAPTER = "Manitec/Plex_Nyhex_v1"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, ADAPTER)
model.eval()

messages = [{"role": "user", "content": "Hello, Nyhex."}]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(
    model.get_input_embeddings().weight.device
)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=128,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

new_tokens = output[0, inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

This example loads the base model plus adapter, not Plex_Nyhex_q4_v1.gguf. Memory requirements depend on precision, device placement, context length, and runtime; the adapter's file size is not the total inference memory requirement. If loading fails, verify package compatibility, available memory, and tokenizer/chat-template compatibility.

GGUF loading path

Download Plex_Nyhex_q4_v1.gguf and load it in a runtime that supports its architecture and GGUF format. Do not pass the GGUF file to PeftModel.from_pretrained or assume the separate PEFT adapter is automatically applied by a GGUF runtime.

Before describing this artifact as a validated standalone Nyhex model, confirm:

• The GGUF architecture and quantization type.
• The source checkpoint and conversion tool/version.
• Whether the Nyhex adapter was merged before conversion.
• The runtime and version used for a successful inference test.
• The chat template and generation settings used for that test.

The filename and file size alone do not establish these details.

Intended use and limitations

This repository provides artifacts for experimental conversational text generation and local integration. No benchmark results or verified claims about persona consistency, reasoning quality, safety, or deployment suitability are documented here.

Generated text may be incorrect, biased, or inappropriate. Independently verify important outputs; do not rely on this model as the sole authority for consequential decisions. The base model's name includes “uncensored”; that name is not a measured safety assessment.

Training and evaluation

The following information has not yet been documented:

• Training dataset names, provenance, consent, and applicable permissions.
• Training method, optimizer, learning rate, epochs or steps, sequence length, and seed.
• Training hardware, software versions, and checkpoint selection.
• Intended Nyhex behavior and its relationship to the earlier Plex Sable naming.
• Evaluation prompts, benchmark results, and known failure cases.
• GGUF conversion and quantization details.

Do not interpret missing information as evidence that a particular training process, dataset, or evaluation was used.

License and attribution

An adapter-specific license has not been established in the inspected repository documentation. Review the base model's license and upstream terms before use, redistribution, or commercial deployment. Training-data permissions and any additional adapter terms also need to be documented. No license identifier is asserted in this model card until those terms are verified.

Documentation scope

This model card replaces the earlier Plex Sable-focused, adapter-only description with repository-specific Plex Nyhex v1 documentation. It documents both artifacts without claiming that the GGUF has been validated or that weights have changed.