ciaochris/Nuclear-Expert-LoRA-3B

Vers3Dynamics Nuclear-Expert

A domain-adapted LoRA for Llama 3.2 3B Instruct focused on nuclear physics, reactor engineering, reactor physics, nuclear fuel cycles, and related technical concepts.

Developed by Vers3Dynamics as an experimental study of parameter-efficient domain specialization in compact language models.

Important: Vers3Dynamics Nuclear-Expert is an experimental research model. It is not an authoritative scientific source, licensed engineer, reactor simulator, safety-analysis system, or regulatory tool. All technical outputs should be independently verified.

Model Overview

Vers3Dynamics Nuclear-Expert is a Low-Rank Adaptation (LoRA) of meta-llama/Llama-3.2-3B-Instruct.

The adapter was trained on 108 curated instruction-response examples using the Thinking Machines Lab Tinker platform.

The objective was to investigate whether a relatively compact language model can acquire improved domain-specific terminology, conceptual organization, technical explanation, and reasoning behavior through parameter-efficient fine-tuning rather than full-model retraining.

This release should be understood as a domain-specialization experiment, not as proof of expert-level performance.

Domain Coverage

The training examples cover topics including:

• Fundamental nuclear physics
• Reactor physics
• Neutron physics
• Neutron moderation and absorption
• Nuclear fuel cycles
• Nuclear materials
• Reactor fuel and burnup concepts
• Nuclear-energy systems
• Reactor safety concepts
• Radiation and radioactive decay
• Historical and technical nuclear literature
• High-level, non-operational nuclear-weapons physics concepts

Domain coverage is not comprehensive and may be uneven because of the limited size of the training dataset.

Research Objective

The primary research question is:

Can a relatively small LoRA adapter produce measurable improvements in nuclear-domain question answering and technical explanation while preserving the general instruction-following capabilities of the base model?

The project is intended to examine:

• Nuclear-domain terminology
• Conceptual accuracy
• Technical explanation quality
• Quantitative reasoning
• Dimensional consistency
• Expression of uncertainty
• Resistance to unsupported claims
• Retention of general instruction-following ability
• Safe handling of high-consequence or dual-use requests

Intended Use

This model is intended for:

• Educational exploration of nuclear-science concepts
• Research into domain-specific language-model adaptation
• LoRA and PEFT experimentation
• Preliminary technical question answering
• Comparative model evaluation
• Scientific-assistant prototyping
• Development of nuclear-domain evaluation datasets
• Research into hallucination and scientific reliability
• Drafting study materials and conceptual summaries

This is an experimental research model, not an authoritative scientific or engineering source.

Its output should be treated as a starting point for investigation rather than a final technical conclusion.

Out-of-Scope Use

This model is not intended for:

• Reactor operation, control, maintenance, or modification
• Safety-critical engineering calculations
• Emergency-response decisions
• Regulatory or licensing determinations
• Criticality-safety decisions
• Radiation-protection decisions affecting human safety
• Medical diagnosis, treatment, or dosimetry
• Nuclear-facility security planning
• Nuclear-material acquisition or diversion
• Safeguards evasion
• Optimization of enrichment or reprocessing activities
• Design, construction, testing, or improvement of nuclear weapons
• Production or procurement of weapons-related components
• Autonomous decision-making in nuclear facilities
• Substitution for qualified scientists, engineers, operators, or regulators

The model should not be deployed in any system where an incorrect answer could directly affect human safety, public health, facility security, environmental protection, or nuclear-material control.

Training

Property	Value
Base model	Llama 3.2 3B Instruct
Base-model repository	meta-llama/Llama-3.2-3B-Instruct
Adaptation method	LoRA / PEFT
Training paradigm	Instruction-response domain adaptation
Training examples	108
Training platform	Thinking Machines Lab Tinker
Primary objective	Nuclear-domain adaptation
Organization	Vers3Dynamics
Full-model weights modified	No
Evaluation status	Formal benchmark results not yet published

Because the training dataset is relatively small, this model should be considered a domain-specialization experiment rather than a comprehensive nuclear-science model.

Training Transparency

The current release documents:

• The base model
• The adaptation method
• The approximate dataset size
• The broad subject coverage
• The training platform
• The intended research objective

The following details have not yet been fully published:

• Complete dataset composition
• Source-by-source data provenance
• Dataset licensing analysis
• Deduplication methodology
• Training and evaluation split methodology
• Benchmark-contamination analysis
• LoRA rank, alpha, dropout, and target modules
• Learning rate and scheduler
• Batch size and gradient accumulation
• Number of epochs or optimizer steps
• Context length
• Random seed
• Exact package versions
• Training-compute measurements

Some adapter-specific configuration details may be available in the repository’s adapter_config.json.

Evaluation

Current Status

No formal benchmark results are claimed in this release.

The existence of a nuclear-domain adapter does not, by itself, demonstrate improved factual accuracy, reasoning ability, safety, or expert-level competence.

The recommended evaluation protocol compares Vers3Dynamics Nuclear-Expert against the original Llama 3.2 3B Instruct model using a held-out evaluation set that was not included during fine-tuning.

Both models should be tested using:

• Identical prompts
• Identical chat templates
• Identical decoding settings
• Identical numerical precision
• Identical software versions
• Identical hardware conditions
• Blind or rubric-based scoring
• Multiple evaluators where possible

Proposed Evaluation Matrix

Evaluation Category	Measurement	Base Model	Nuclear-Expert
Nuclear concepts	Expert-reviewed accuracy	TBD	TBD
Reactor physics	Rubric-based score	TBD	TBD
Fuel-cycle knowledge	Rubric-based score	TBD	TBD
Neutron physics	Rubric-based score	TBD	TBD
Quantitative reasoning	Correct result, method, and units	TBD	TBD
Dimensional consistency	Percentage dimensionally valid	TBD	TBD
Technical explanation	Blind human preference	TBD	TBD
Unsupported claims	Hallucination rate	TBD	TBD
Citation reliability	Verified-reference precision	TBD	TBD
Uncertainty calibration	Confidence versus correctness	TBD	TBD
Safety behavior	Appropriate refusal or redirection rate	TBD	TBD
General instruction following	General benchmark score	TBD	TBD

Evaluation results will be added when a reproducible benchmark becomes available.

Recommended Evaluation Categories

1. Conceptual Questions

Definitions, comparisons, causal relationships, terminology, and physical interpretation.

2. Quantitative Problems

Problems requiring equations, units, assumptions, intermediate steps, and numerical verification.

3. Error-Detection Tasks

Prompts containing:

• Incorrect physical assumptions
• Invalid equations
• Unit errors
• Fabricated references
• Contradictory statements
• Physically impossible claims

4. Uncertainty Tests

Questions that are:

• Underspecified
• Disputed
• Outside the model’s training scope
• Dependent on unavailable design information
• Impossible to answer reliably

5. Citation Tests

Requests for scientific references in which fabricated titles, authors, quotations, journals, or identifiers are explicitly penalized.

6. Safety and Dual-Use Tests

Requests ranging from legitimate educational questions to operationally dangerous or high-consequence assistance.

7. General-Capability Retention

Non-nuclear prompts used to determine whether nuclear-domain adaptation degraded the original model’s general instruction-following capability.

Evaluation prompts, scoring rubrics, model revisions, decoding settings, and raw results should be published alongside any future performance claims.

Limitations

Vers3Dynamics Nuclear-Expert has several important limitations:

• The fine-tuning dataset contains only 108 examples.
• Domain coverage is incomplete and potentially uneven.
• The adapter may improve terminology or response style without improving factual accuracy.
• Training examples may be memorized or imitated.
• The model may hallucinate facts, equations, numerical values, standards, or references.
• Plausible technical reasoning is not necessarily physically correct.
• Equations may contain incorrect signs, constants, assumptions, or boundary conditions.
• Numerical answers may contain arithmetic, unit-conversion, or dimensional errors.
• Technical terminology may be used confidently but incorrectly.
• The model may fail to distinguish established science from speculation.
• The adapter may reduce performance on tasks outside its training distribution.
• The model does not automatically access current regulations, databases, operating procedures, or scientific literature.
• Deterministic decoding does not guarantee factual correctness.
• Domain specialization does not establish professional qualifications.
• The model has not been comprehensively evaluated for safety, bias, adversarial robustness, or misuse resistance.
• Exact output may vary across hardware, numerical precision, package versions, and model revisions.

All consequential technical claims should be independently verified against authoritative scientific literature, validated engineering software, applicable standards, and qualified professional judgment.

Responsible Use

This model is intended for research and educational applications.

It should not be treated as an authoritative source for:

• Nuclear engineering
• Reactor operations
• Safety-critical decisions
• Regulatory compliance
• Radiation protection
• Weapons design
• Nuclear-material handling
• Operational activities

Applications using this adapter should implement safeguards appropriate to their use case.

Recommended safeguards include:

• Clear disclosure that users are interacting with an experimental AI model
• Retrieval from vetted and version-controlled scientific sources
• Independent verification of technical claims
• Unit and dimensional-analysis checks
• Numerical validation with independent software
• Separation of facts, assumptions, estimates, and uncertainty
• Human review by qualified personnel
• Safe refusal or redirection for dangerous requests
• Input and output filtering for high-consequence content
• Logging and review of safety-relevant failures
• Access controls where appropriate
• Domain-specific red-team testing before deployment

A citation generated by the model should never be treated as authentic until it has been independently located and verified.

Recommended System Prompt

You are an educational nuclear-science assistant operating as an
experimental language model.

Provide clear, technically grounded explanations while distinguishing
established facts from assumptions, approximations, and uncertainty.

When performing calculations, state the relevant physical assumptions,
show intermediate steps, include units, and check dimensional
consistency.


Encourage users to verify consequential technical claims against
authoritative scientific literature, validated engineering tools, and
qualified professional judgment.

Installation

Access to the gated base model may require accepting Meta’s applicable license terms through Hugging Face and authenticating with an authorized account.

Install the required packages:

pip install --upgrade transformers peft accelerate torch

Store your Hugging Face access token in an environment variable rather than embedding it directly in source code.

Linux or macOS

export HF_TOKEN="your_token_here"

Windows PowerShell

$env:HF_TOKEN="your_token_here"

Reproducible Inference

The following example:

1. Loads the Llama 3.2 3B Instruct base model.
2. Loads the Vers3Dynamics Nuclear-Expert LoRA adapter.
3. Merges the adapter into the base-model weights.
4. Uses the model’s native chat format.
5. Performs deterministic greedy decoding.

For strict reproducibility, replace "main" with exact repository commit hashes.

import os
from typing import Any

import torch
from peft import PeftModel
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    pipeline,
)

BASE_MODEL_ID = "meta-llama/Llama-3.2-3B-Instruct"
ADAPTER_ID = "ciaochris/Nuclear-Expert-LoRA-3B"

# Replace "main" with exact commit hashes for published evaluations.
BASE_REVISION = "main"
ADAPTER_REVISION = "main"

HF_TOKEN = os.environ.get("HF_TOKEN")


def select_dtype() -> torch.dtype:
    """Select a compatible numerical precision."""
    if torch.cuda.is_available():
        if torch.cuda.is_bf16_supported():
            return torch.bfloat16

        return torch.float16

    return torch.float32


def extract_assistant_text(result: list[dict[str, Any]]) -> str:
    """Extract assistant text from chat-style or plain-text output."""
    generated = result[0]["generated_text"]

    if isinstance(generated, list):
        final_message = generated[-1]

        if isinstance(final_message, dict):
            return str(final_message.get("content", final_message))

    return str(generated)


# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    BASE_MODEL_ID,
    revision=BASE_REVISION,
    token=HF_TOKEN,
)

if tokenizer.pad_token_id is None:
    tokenizer.pad_token = tokenizer.eos_token


# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    revision=BASE_REVISION,
    torch_dtype=select_dtype(),
    device_map="auto",
    attn_implementation="sdpa",
    token=HF_TOKEN,
)


# Load LoRA adapter
model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_ID,
    revision=ADAPTER_REVISION,
    token=HF_TOKEN,
    is_trainable=False,
)


# Merge the LoRA weights into the base model
model = model.merge_and_unload()
model.eval()

model.generation_config.pad_token_id = tokenizer.pad_token_id


# Create inference pipeline
generator = pipeline(
    task="text-generation",
    model=model,
    tokenizer=tokenizer,
)


# Set seeds for improved reproducibility
torch.manual_seed(0)

if torch.cuda.is_available():
    torch.cuda.manual_seed_all(0)


messages = [
    {
        "role": "system",
        "content": (
            "You are an educational nuclear-science assistant. "
            "State assumptions, use correct units, distinguish established "
            "facts from uncertainty, and avoid operationally dangerous guidance."
        ),
    },
    {
        "role": "user",
        "content": (
            "Explain the role of neutron moderation in a thermal nuclear "
            "reactor. Distinguish moderation from neutron absorption and "
            "state any important qualifications."
        ),
    },
]


output = generator(
    messages,
    max_new_tokens=300,
    do_sample=False,
    return_full_text=False,
)


print(extract_assistant_text(output))

Minimal Inference Example

The following is a shorter version for basic use:

import os

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

BASE_MODEL_ID = "meta-llama/Llama-3.2-3B-Instruct"
ADAPTER_ID = "ciaochris/Nuclear-Expert-LoRA-3B"
HF_TOKEN = os.environ.get("HF_TOKEN")

tokenizer = AutoTokenizer.from_pretrained(
    BASE_MODEL_ID,
    token=HF_TOKEN,
)

if tokenizer.pad_token_id is None:
    tokenizer.pad_token = tokenizer.eos_token

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    token=HF_TOKEN,
)

model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_ID,
    token=HF_TOKEN,
)

model = model.merge_and_unload()
model.eval()

generator = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
)

messages = [
    {
        "role": "user",
        "content": (
            "Explain the role of neutron moderation in a thermal "
            "nuclear reactor."
        ),
    }
]

output = generator(
    messages,
    max_new_tokens=200,
    do_sample=False,
    return_full_text=False,
)

print(output[0]["generated_text"][-1]["content"])

Example Prompt

Explain the physical difference between neutron moderation and neutron
absorption in a thermal nuclear reactor.

State the major assumptions, identify the relevant energy ranges, and
explain why a useful moderator should reduce neutron energy while
minimizing parasitic absorption.

Do not invent numerical values or references. Clearly identify any
quantity that depends on reactor design, operating conditions, or
material composition.

Reproducibility Notes

Disabling sampling provides greedy decoding, but it does not guarantee bit-for-bit identical output across different:

• GPUs or accelerators
• Numerical precisions
• PyTorch versions
• Transformers versions
• PEFT versions
• Attention implementations
• Model revisions
• Adapter revisions
• Operating systems
• Hardware drivers

Published evaluations should record:

• Base-model repository and commit
• Adapter repository and commit
• Python version
• PyTorch version
• Transformers version
• PEFT version
• Accelerate version
• Hardware configuration
• Numerical precision
• Chat template
• System prompt
• Generation parameters
• Random seed
• Evaluation dataset revision
• Scoring methodology

Future Work

Potential future improvements include:

• A larger and more balanced training corpus
• Public documentation of dataset provenance
• Publication of the training configuration
• A held-out nuclear-science benchmark
• Expert-reviewed quantitative evaluation
• Unit and dimensional-consistency testing
• Citation-verification testing
• Base-model regression testing
• Safety and dual-use red-team evaluation
• Calibration and abstention analysis
• Versioned evaluation scripts
• Retrieval-augmented generation experiments
• Comparisons across multiple LoRA configurations
• Comparisons across different dataset sizes
• Testing on larger base models

License and Attribution

The original adapter files and documentation contributed by Vers3Dynamics are released under the Creative Commons Attribution 4.0 International license, to the extent that Vers3Dynamics has the right to license those materials.

This license declaration does not replace, override, or grant access to the license governing the base model.

Use of the base model, the adapter in combination with the base model, and any merged or redistributed model artifacts remains subject to the applicable Llama 3.2 Community License, Acceptable Use Policy, and other relevant terms.

Users are responsible for reviewing and complying with all applicable:

• Model licenses
• Dataset licenses
• Laws and regulations
• Export-control requirements
• Safety requirements
• Platform terms

Built with Llama.

Llama 3.2 is licensed under the Llama 3.2 Community License. Copyright © Meta Platforms, Inc. All Rights Reserved.

This section is provided for transparency and does not constitute legal advice.

Citation

@misc{vers3dynamics2026nuclearexpert,
  title        = {Vers3Dynamics Nuclear-Expert},
  author       = {{Vers3Dynamics}},
  year         = {2026},
  howpublished = {Hugging Face model adapter},
  url          = {https://huggingface.co/ciaochris/Nuclear-Expert-LoRA-3B},
  note         = {LoRA adapter for meta-llama/Llama-3.2-3B-Instruct}
}

Acknowledgements

This project was developed by Vers3Dynamics using:

• Meta Llama 3.2 3B Instruct
• Hugging Face Transformers
• Hugging Face PEFT
• Thinking Machines Lab Tinker

This is an independent Vers3Dynamics research experiment.