You can deploy this model and start earning money today!

Presence-Beyond-Form/
gemma-4-31b-rv-lora-v1

Name: gemma-4-31b-rv-lora-v1
Availability: InStock

Presence-Beyond-Form/RV_trening_AI

google/gemma-4-31B-it

Views

No views yet

Gemma 4 31B RV LoRA v1

Gemma 4 31B RV LoRA v1 is a PEFT LoRA adapter trained on top of google/gemma-4-31B-it.

The adapter was created for experimental work on structured AI Remote Viewing (RV) protocols developed by Presence Beyond Form.

Important: this repository contains the LoRA adapter only.
It does not contain the full Gemma 4 31B model and it is not a merged checkpoint.

Base model

Base model: google/gemma-4-31B-it
Adapter: Presence-Beyond-Form/gemma-4-31b-rv-lora-v1
Method: LoRA / PEFT
Training objective: Supervised Fine-Tuning (SFT)
Base-model precision during training: BF16
Quantization during training: none
Merged with base model: no

The LoRA adapter was applied only to the language-model transformer. Gemma 4's vision components were not fine-tuned.

Training Dataset

The training data used for this adapter is publicly available on Hugging Face:

Dataset: Presence-Beyond-Form/RV_trening_AI

Training data version: dataset_1_0

The Gemma 4 31B RV LoRA v1 training run used the following three SFT files from this dataset version:

File	Examples
`datasetV1_sft_1_0.jsonl`	411
`datasetV2_stf_1_0.jsonl`	244
`datasetV3_sft_1_0.jsonl`	86
Total	741

Each record uses a JSONL text field containing the supervised fine-tuning example.

The dataset is published separately so that the training material can be inspected independently and the fine-tuning procedure can be reproduced.

Training configuration

LoRA

Parameter	Value
LoRA rank (`r`)	16
LoRA alpha	32
LoRA dropout	0.05
Bias	none
Trainable parameters	122,429,440
Total parameters reported during training	31,395,515,952
Trainable share	0.3900%

The adapter targeted the following projection modules inside the language model:

q_proj
k_proj
v_proj
o_proj
gate_proj
up_proj
down_proj

SFT hyperparameters

Parameter	Value
Epochs	3
Per-device batch size	1
Gradient accumulation	8
Effective batch size	8
Learning rate	`1e-5`
Scheduler	linear
Warmup steps	28
Optimizer	`adamw_torch`
Maximum sequence length	1024
Gradient checkpointing	enabled
Precision	BF16
Quantization	none
Packing	disabled
Seed	42

Training environment

Hardware: NVIDIA A100 80 GB
Training service: Hugging Face Jobs
Transformers: 5.15.0
TRL: 1.9.2
PEFT: 0.20.0

Training results

The successful run completed all 3 epochs.

Metric	Result
Optimizer steps	279
Training runtime	2376.43 s
Training runtime	~39 min 36 s
Training samples / second	0.935
Training steps / second	0.117
Total training tokens	191,814
Reported training loss	4.9099
Mean token accuracy	0.5066
Reported entropy	2.4874

During the run, the logged step loss decreased from approximately 15.4 at the beginning to approximately 2.6-2.7 near the end of the third epoch.

These are training metrics only. They are not held-out evaluation results.

Intended use

This adapter is intended for research and experimentation involving structured AI Remote Viewing prompting and related protocol-following behavior.

Potential uses include:

controlled RV research sessions;
comparison of the original Gemma 4 31B IT model against the LoRA-adapted version;
experiments on instruction adherence, response structure, vocabulary, and protocol stability;
integration into research harnesses or inference servers that support PEFT/LoRA adapters.

This model card does not claim that remote viewing is scientifically established or that an AI system can acquire information through anomalous means. Such claims require independent experimental evidence.

How to use

Install the required packages:

pip install -U transformers peft accelerate torch

Load the original Gemma 4 31B IT model and attach the adapter:

python

import torch
from transformers import AutoTokenizer, Gemma4ForConditionalGeneration
from peft import PeftModel

base_model_id = "google/gemma-4-31B-it"
adapter_id = "Presence-Beyond-Form/gemma-4-31b-rv-lora-v1"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)

base_model = Gemma4ForConditionalGeneration.from_pretrained(
    base_model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

model = PeftModel.from_pretrained(
    base_model,
    adapter_id,
)

model.eval()

prompt = "Your prompt here."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
    )

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

For the closest reproduction of the training setup, use the non-quantized BF16 base model.

Required model combination

To use this adapter you need both repositories:

text

google/gemma-4-31B-it
+
Presence-Beyond-Form/gemma-4-31b-rv-lora-v1

The adapter may be loaded dynamically with PEFT or served by an inference engine that supports LoRA adapters.

Evaluation status

No separate held-out evaluation dataset was used during fine-tuning.

The adapter should therefore be evaluated independently against the unmodified google/gemma-4-31B-it model before drawing conclusions about:

instruction adherence;
response structure;
vocabulary behavior;
stability;
output quality;
performance in RV research sessions.

For controlled comparisons, use the same prompts, generation parameters, hardware/provider configuration, and scoring method for both the base model and the adapted model.

Limitations

The SFT dataset contains 741 examples.
Training metrics are not independent evaluation metrics.
The adapter may reproduce patterns or wording from its training material.
Behavior can vary with prompt structure, sampling parameters, context length, and inference implementation.
Quantized inference may behave differently from the BF16 setup used during training.
The adapter does not establish the scientific validity of remote viewing or anomalous information acquisition.

License and base-model terms

This repository contains an adapter derived from google/gemma-4-31B-it.

The base model is published by Google under the license and usage terms shown on the original Gemma 4 model repository. Users should review and comply with those terms when using the base model together with this adapter.

Acknowledgements

Google DeepMind for Gemma 4.
Hugging Face for Transformers, TRL, PEFT, Hub, and Jobs.
Presence Beyond Form for the research workflow, dataset preparation, training setup, and experimental design.

Model card

Prepared by Presence Beyond Form for the AI Remote Viewing research project.