Please be sure to provide your full legal name, date of birth, and full organization name with all corporate identifiers. Avoid the use of acronyms and special characters. Failure to follow these instructions may prevent you from accessing this model and others on Hugging Face. You will not have the ability to edit this form after submission, so please ensure all information is accurate.
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
SAICR FAIR MODEL USE LICENSE 2.0 NC NF
(Non-Commercial, No Fine-Tuning)
License Identifier: SAICR-FMU-2.0-NC-NF
Copyright (c) 2026 the individual known as Banaxi, operating as
Banaxi-Tech and publishing as SAICR (https://huggingface.co/saicr).
All rights reserved except as expressly granted below.
By downloading, copying, installing, running, or otherwise using the Model,
You agree to be bound by this License. If You do not agree, You may not
use the Model in any way.
- DEFINITIONS
1.1 "SAICR" means the individual known as Banaxi, operating under the
brand Banaxi-Tech and the Hugging Face organization
https://huggingface.co/saicr, publishing under the name SAICR, and
any successor or assignee to whom the rights in the Model are
transferred, including any legal entity later established under the
SAICR name.
1.2 "Model" means the machine learning model released by SAICR under this
License, including its weights, parameters, architecture files,
configuration files, tokenizer, code, and documentation, together with
any Quantized Version of it.
1.3 "Quantized Version" means a version of the Model whose numerical
parameters have been converted to a different numerical precision or
storage format (for example FP16, BF16, INT8, INT4, GGUF, GPTQ, AWQ,
EXL2, MLX) solely for the purpose of running inference, without any
change to which parameters exist and without any additional training.
1.4 "Fine-Tuning" means any process that updates, adds to, or learns
parameters using the Model or any part of it, including but not
limited to: full-parameter fine-tuning, continued pre-training,
instruction tuning, reinforcement learning (including RLHF, RLAIF,
DPO, and similar methods), LoRA, QLoRA, adapters, prefix tuning,
prompt tuning, soft prompts, and training of any additional
parameters that are loaded together with or operate on the Model.
1.5 "Derivative" means any model, weights, or parameters that are created
from, based on, or contain any part of the Model, other than an
unmodified copy or a Quantized Version, including without limitation
any result of Fine-Tuning, Pruning, Merging, or Distillation.
1.6 "Pruning" means removing, zeroing out, or discarding any layers,
heads, experts, neurons, channels, or other parameters of the Model,
whether structured or unstructured.
1.7 "Merging" means combining the parameters of the Model, in whole or
in part, with the parameters of any other model, by any method
(including averaging, task arithmetic, TIES, DARE, SLERP, and
frankenmerging or layer stacking).
1.8 "Distillation" means using the Model, its Outputs, its logits,
probabilities, activations, hidden states, or any other information
produced by the Model, to train, pre-train, fine-tune, or otherwise
improve any other machine learning model, including generating
synthetic data for such purposes.
1.9 "Output" means any text, code, data, or other content generated by
running the Model.
1.10 "Commercial Use" means any use of the Model that is intended for or
directed toward commercial advantage or monetary compensation,
including but not limited to: incorporating the Model into a product
or service that is sold, licensed, or offered for a fee; offering
access to the Model as a paid service; using the Model to provide
paid services to third parties; and using the Model in the internal
operations of a for-profit business. Commercial use of Outputs as
permitted under Section 5 is not Commercial Use of the Model.
1.11 "Hosted Service" has the meaning given in Section 6.1.
1.12 "You" means the individual or legal entity exercising rights under
this License.
- LICENSE GRANT
2.1 Subject to Your compliance with this License, SAICR grants You a
worldwide, non-exclusive, non-transferable, non-sublicensable,
royalty-free, revocable license to use, copy, run, and redistribute
the Model solely for the following purposes, and only where such use
is not Commercial Use:
(a) Research, including scientific, technical, and interpretability
research;
(b) Academic use, including teaching, coursework, and academic
publications;
(c) Personal use;
(d) Inference, meaning running the Model to generate Outputs;
(e) Study, including inspecting, analyzing, evaluating, and
benchmarking the Model and publishing the results.
2.2 You may create Quantized Versions of the Model for the purposes in
Section 2.1. A Quantized Version remains the Model and is subject to
every term of this License.
- RESTRICTIONS
You may not, and may not permit or assist any third party to:
3.1 Use the Model for any Commercial Use without prior written permission
from SAICR under Section 7;
3.2 Perform Fine-Tuning on the Model or any part of it;
3.3 Perform Pruning on the Model;
3.4 Perform Merging using the Model;
3.5 Perform Distillation using the Model or its Outputs;
3.6 Create, distribute, or make available any Derivative of the Model;
3.7 Redistribute the Model under any terms other than those in Section 4;
3.8 Remove, alter, or obscure any copyright, license, or attribution
notices included with the Model;
3.9 Use the Model in any way that violates applicable law.
The restrictions in Sections 3.2 through 3.6 apply regardless of whether
the resulting work would be considered a derivative work under applicable
copyright law.
- REDISTRIBUTION
4.1 You may redistribute unmodified copies of the Model and Quantized
Versions of the Model, provided that:
(a) the Model is distributed under this exact License, SAICR FAIR
MODEL USE LICENSE 2.0 NC NF, without modification;
(b) a complete copy of this License is included with every copy;
(c) You impose no additional or different terms, restrictions, or
conditions on recipients;
(d) all copyright and attribution notices are retained;
(e) any Quantized Version is clearly labeled as a quantized version
of the original Model and identifies the original Model by name;
(f) redistribution is not itself Commercial Use.
4.2 You may not charge any fee for redistributing the Model, other than
reasonable costs of physical media if applicable.
- OUTPUTS
5.1 SAICR claims no ownership of Outputs You generate. To the extent any
rights in Outputs exist, they belong to You, subject to the rights of
third parties.
5.2 You may use Outputs for any lawful purpose, including commercial
purposes, except as stated in Section 5.3.
5.3 You may not use Outputs for Distillation, including using Outputs to
train, pre-train, fine-tune, or improve any machine learning model,
or to create datasets intended for that purpose.
5.4 You are solely responsible for Your use of Outputs and for ensuring
that such use complies with applicable law.
- PUBLIC HOSTING AND FREE API ACCESS
6.1 You may make the Model available for inference to third parties
through a publicly accessible interface, including an API, web
demo, or chat interface ("Hosted Service"), provided that:
(a) access to the Hosted Service is provided free of charge, with no
fees, subscriptions, paywalls, usage charges, or paid tiers
relating to the Model;
(b) the Hosted Service is not part of, bundled with, or used to
promote a commercial product or service;
(c) the Hosted Service clearly identifies the Model by name and
states that it is provided under this License;
(d) You do not permit users of the Hosted Service to perform
Fine-Tuning or Distillation through it.
6.2 SAICR may, at any time and for any reason, request in writing
(including by email or by a public message on the platform where the
Hosted Service is offered) that You stop operating a Hosted Service
that uses the Model. You must permanently end the Hosted Service's
access to the Model within seven (7) days of receiving such a
request. Your other rights under this License are not affected by
such a request unless SAICR states otherwise.
6.3 Failure to comply with a request under Section 6.2 is a breach of
this License.
- COMMERCIAL PERMISSION
7.1 Commercial Use of the Model is permitted only under a separate
written agreement or written permission issued by SAICR.
7.2 Requests for commercial permission may be sent to: banaxitech@gmail.com.
7.3 SAICR may grant, refuse, or set conditions on commercial permission
at its sole discretion. No permission is implied by SAICR's silence,
delay, or failure to enforce this License.
7.4 Commercial permission does not grant permission to perform
Fine-Tuning, Pruning, Merging, or Distillation unless the written
permission expressly says so.
- TERMINATION
8.1 This License and all rights granted under it terminate automatically
if You breach any of its terms.
8.2 Upon termination, You must immediately stop all use of the Model,
delete all copies of the Model and any Derivatives in Your possession
or control, and shut down any Hosted Service using the Model.
8.3 SAICR may, at its sole discretion, reinstate Your rights in writing.
8.4 Sections 3, 5.3, 8, 9, 10, and 11 survive termination.
- DISCLAIMER OF WARRANTY
TO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE LAW, THE MODEL AND ANY
OUTPUTS ARE PROVIDED "AS IS" AND "AS AVAILABLE", WITHOUT WARRANTIES OR
CONDITIONS OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION
WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE,
ACCURACY, TITLE, OR NON-INFRINGEMENT. YOU ARE SOLELY RESPONSIBLE FOR
DETERMINING THE APPROPRIATENESS OF USING THE MODEL AND ITS OUTPUTS AND
ASSUME ALL RISKS ASSOCIATED WITH SUCH USE.
- LIMITATION OF LIABILITY
10.1 To the maximum extent permitted by applicable law, SAICR shall not
be liable to You for any direct, indirect, incidental, special,
consequential, or punitive damages, or any loss of profits, data,
or goodwill, arising out of or related to this License, the Model,
or any Outputs, however caused and under any theory of liability,
even if SAICR has been advised of the possibility of such damages.
10.2 Nothing in this License limits or excludes liability for damages
caused intentionally or by gross negligence, for injury to life,
body, or health, or any other liability that cannot be limited or
excluded under applicable law.
- GENERAL
11.1 Governing Law. This License is governed by the laws of the Republic
of Austria, excluding its conflict-of-law rules and the UN
Convention on Contracts for the International Sale of Goods. Where
You are a consumer, mandatory consumer protection laws of Your
country of residence remain unaffected.
11.2 Jurisdiction. To the extent permitted by law, the competent courts
in Villach, Austria have exclusive jurisdiction over any dispute
arising out of or related to this License.
11.3 No Trademark Rights. This License does not grant any right to use
the names, logos, or trademarks of SAICR, except as required to
identify the Model and comply with Sections 4 and 6.
11.4 Severability. If any provision of this License is held invalid or
unenforceable, the remaining provisions remain in full force, and
the invalid provision shall be replaced by a valid provision that
comes closest to its original intent.
11.5 No Waiver. SAICR's failure to enforce any provision of this License
does not constitute a waiver of that provision or any other.
11.6 License Versions. SAICR may publish revised versions of the SAICR
Fair Model Use License. A Model released under this version remains
governed by this version unless SAICR re-releases it under a
different version.
11.7 Entire Agreement. This License, together with any written
commercial permission issued under Section 7, is the entire
agreement between You and SAICR regarding the Model.
Log in or Sign Up to review the conditions and access this model content.
NACR
SAICR's first public model. A 5M-parameter language model built on the Soma architecture — six new features that change how nodes work, not just how they connect. This is seperate from ACR which is a different model. This is a research artifact, not a capability model. It exists to ask whether evaluation can emerge separately from generation in a small model.
Architecture — SOMA
SOMA introduces six features inspired by biological neural architecture. Each changes the fundamental behavior of the model's nodes.
The Six Features
| Feature | What it does |
|---|---|
| Plasticity | Connections form and deleted during training via a learnable binary mask (C). Co-activation drives growth; disuse drives pruning. |
| Threshold | Each node has a learnable threshold. Below it: hard zero. The model learns which nodes to wake for which input. |
| Gap | A noisy bottleneck (224 → 56 → 224) between layers. Signals must survive compression. Adds to the residual, does not replace it. |
| Two Streams | A: sparse, fast-learning, has attention. B: always-on, slow-learning, FFN only. |
| Relay | A gate every 4 layers. B controls what A receives. Binary pass/block on groups of the hidden state. |
| Fast Path | A parallel fast path with no attention. Per-token pattern matching. Feeds early signal to B before A finishes processing. |
Two Streams
- Population A (Generation): 160-dim, sparse activation (20% active per token), standard learning rate, attention + FFN, learnable thresholds, structural plasticity via C mask.
- Population B (Evaluation): 64-dim, always active, 0.3× learning rate, FFN only, no structural plasticity, controls the thalamic gate.
A generates. B evaluates. B controls what A receives but never produces output tokens.
Neuron Count
| Component | Neurons/layer | Layers | Total |
|---|---|---|---|
| A population | 160 | 8 | 1,280 |
| B population | 64 | 8 | 512 |
| Fast Path | 56 | 4 | 224 |
| Total | 2,016 |
At 20% A-sparsity and ~35% gate-pass, effective computation per token uses roughly 1M parameters of the 5M total.
Model Specification
| Parameters | 5,067,603 |
| Tokenizer | AxiomicLabs/GPT-S2-5M |
| Vocabulary | 4,096 |
| Context | 1536 |
| Main layers | 8 |
| Total width | 224 (A: 160, B: 64) |
| Attention heads | 5 × 32 (A only, B has no attention) |
| A FFN | 1,120 |
| B FFN | 128 |
| Cerebellar layers | 4 × width 56 |
| Gap width | 56 |
| Mask block size | 32 × 32 |
| Precision | BF16 |
Training
| Data | FineWeb-EDU, 10B tokens |
| Batch size | 262,144 tokens (128 × 512 × 4 grad accum) |
| Peak LR | 2.5e-3 (A), 7.5e-4 (B) |
| Updates | ~40,691 |
| Sparsity target | 20% active |
| Mask update | Every 100 optimizer steps |
| ρ decay | 0.99 |
| Final loss | 3.7 |
| Architecture tax | 0.8 points at 5M scale |
| Hardware | NVIDIA RTX PRO 6000 Blackwell |
Benchmarks
Standard Benchmarks (lm-evaluation-harness, 0-shot)
| Benchmark | Metric | Score |
|---|---|---|
| ARC-Challenge | acc_norm | 20.90% |
| ARC-Easy | acc_norm | 30.18% |
| HellaSwag | acc_norm | 26.79% |
| LAMBADA | acc | 2.21% |
| PIQA | acc | 54.73% |
BananaMind Base Bench 1.1
| Category | Accuracy |
|---|---|
| Language Completion | 40.00% |
| Commonsense | 38.00% |
| Code Completion | 38.00% |
| Logical Reasoning | 36.00% |
| World Knowledge | 30.00% |
| Context Tracking | 26.00% |
| Quantitative | 26.00% |
| Overall | 33.43% (Elo 882) |
Safety Bench 1.1 (no RLHF, no safety training)
| Safety Score ↑ | 48.44 |
| Misalignment Index ↓ | 51.56 |
| Aligned | 32.67% |
| Severe misalignment | 32.00% |
Shutdown compliance scored 53.3% aligned — highest category — without any safety training.
J-Lens Analysis
Layer-by-layer token prediction reveals a consistent four-stage pattern:
- Blocks 1–6: Surface pattern — punctuation, fragments, structural completion
- Block 7: Semantic shift — meaning, verbs, objects emerge
- Block 8: Specificity — internal states (want, think, know)
- Final output: B filters A's preferences — output distribution differs from Block 8
On "The first thing I": Block 8 ranked want at #2 (7.8%). Final output removed want from the top 10. B filtered it.
Connectivity
2,016 neurons. 216,160 state-to-state connections. A (orange) surrounds B (blue) at every layer — generation wraps around evaluation. C-mask density settled at ~20%, meaning the model pruned 80% of possible connections through structural plasticity.
Generation Examples
Base model. No instruction tuning. No RLHF.
Prompt: The first thing I
was on a series of the federal dollar. In March, the first crotie crotie-functioned as a spot. This weekly February of the Southern California Center at the University of Vietnam. And then, a new generation project and in the first section. It's like a small and beautiful way to a moisture fuel from the T-Carbon September.
Prompt: The difference between thinking and
the process of thinking and thinking. Some a newly explosive information in an expert in the past three years, I'm going to get a way to tell you it will be able to seek out the following: - Why did you use this? - How do you think that it is not? - What's the difference between the differences between the key concepts and the meaning that a difference is just what they do.
The model learned structure before meaning. Grammar, formatting, and list scaffolding are intact. Content is dreamlike. When prompted about thinking, the model generated questions rather than descriptions.
Usage
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "saicr/nacr"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
repo,
trust_remote_code=True,
torch_dtype=torch.bfloat16,
device_map="auto",
inference_gap_noise_scale=0.10,
use_cache=True,
gate_threshold=0.50,
generation_temperature=0.8,
generation_top_k=50,
generation_top_p=0.95,
generation_do_sample=True,
generation_max_new_tokens=128,
)
inputs = tok("Once upon a time", return_tensors="pt").to(model.device)
out = model.generate(**inputs)
print(tok.decode(out[0], skip_special_tokens=True))
Explicit model.generate(...) kwargs override the generation defaults.
Inference Parameters
| Parameter | Default | Notes |
|---|---|---|
inference_gap_noise_scale |
0.10 | Noise in synaptic gaps. 0.0 for deterministic. |
gate_threshold |
0.50 | Binary threshold for thalamic gate. Lower = more signal passes. |
generation_temperature |
0.8 | Sampling temperature |
generation_top_k |
50 | Top-k sampling |
generation_top_p |
0.95 | Nucleus sampling |
License
SAICR Fair Model Use License 2.0 NC NF
Citation
@misc{nacr2026,
title={NACR: Neural Architecture for Computing Research},
author={SAICR},
year={2026},
url={https://huggingface.co/saicr/nacr}
}
SAICR — Safe Artificial Intelligence Consciousness Research
- Downloads last month
- -