Hancock1111/ThermalGuardV_v1_1

Model description

ThermalGuardV-v1_1 is a Vision-Language Model (VLM) based on Qwen2.5-VL-7B-Instruct, fine-tuned with LoRA (Low-Rank Adaptation) and fully merged for direct deployment. It excels in materials science domains, particularly:

• Thermal Barrier Coatings (TBCs)

• High-Entropy Alloys (HEAs)

• High-Temperature Oxidation

This model has been enhanced with technical knowledge about advanced materials for high-temperature applications, including composition design, microstructure characterization, performance evaluation, and failure mechanisms.

Model Variants

• Full Precision merged model (fp16)

Intended uses & limitations

Intended Uses

• Multimodal Analysis: Process and interpret both text and visual data (e.g., microstructure images, phase diagrams, SEM/TEM micrographs).
• Technical documentation generation for high-temperature materials
• Research assistance in materials science
• Answering technical questions about TBCs, HEAs, and oxidation behavior
• Literature review support for materials engineering
• Educational tool for materials science students

Limitations

• Domain Specificity: The model performs best on topics related to thermal barrier coatings, high-entropy alloys, and high-temperature oxidation. Performance on unrelated materials science topics may vary.
• Visual Understanding Constraints:

  1. May struggle with extremely low-resolution or noisy images.

  2. Limited ability to quantify microstructural features (e.g., exact grain size distribution) without calibration.

  3. Interpretations of complex multi-phase microstructures may require validation.

• Technical Accuracy: Although fine-tuned with domain knowledge, the model's outputs should be verified against authoritative sources for critical applications.
• Numerical Precision: The model may approximate numerical values or properties—exact figures should always be confirmed experimentally.
• Knowledge Cutoff: The model's knowledge is current only up to its training data cutoff date.
• Safety-Critical Applications: This model should not be used as the sole decision-making tool for safety-critical applications without human oversight.
• This model is only optimized for Chinese (中文)

Training data

The model was fine-tuned on a combination of:

1. Technical literature about:

• Thermal barrier coatings (YSZ, gadolinium zirconate, etc.)

• High-entropy alloy systems (CoCrFeMnNi, refractory HEAs, etc.)

• High-temperature oxidation mechanisms

2. Curated datasets:

• ShareGPT conversations with materials science focus

Metric	Value
Total Conversations	18,284
Avg. Turns per Conv.	2.00
Max Turns	2
Avg. Chars per Turn	191.98
User Turns	18,284
Assistant Turns	18,284

17,473,536 input tokens

• Keyword cloud

Training procedure

Training hyperparameters

• Base Model: Qwen2.5-VL-7B-Instruct

• Fine-tuning Method: LoRA (Low-Rank Adaptation), later fully merged

• Learning Rate: 0.0001

• Batch Size: 4 (effective size 8 with gradient accumulation)

• Epochs: 3

• Optimizer: AdamW (β₁=0.9, β₂=0.999, ε=1e-08)

• Scheduler: Cosine learning rate schedule

• Mixed Precision: Native AMP

Framework versions

• PEFT 0.15.2
• Transformers 4.53.3
• Pytorch 2.5.1+cu124
• Datasets 3.6.0
• Tokenizers 0.21.1

Baseline Model Performance

General performance

Dataset	Deepseek-R1	Qwen3-8B	Qwen2.5-VL-7B	ThermalGuard_v1_4	ThermalGuardV_v1_1
CEVAL (Accuracy)	/	0.532	0.726	0.731	0.696
GPQA Diamond (Pass@1)	71.5 +	31.3	28.5	41.9	30.8
Simple-Materials-QA (Rouge-1-R)	0.419	0.497	0.262	0.479	0.453
MMMU_VAL	/	/	0.454	/	0.467
DocVQA	/	/	92.4	/	92.1
OCRVQA	/	/	70.3	/	69.7
TextVQA	/	/	84.6	/	84.4
MMStar	/	/	0.608	/	0.608
+The scores are from https://api-docs.deepseek.com/news/news250120s

Professional knowledge performance

The dataset was used to evaluate several recent models with the following qualitative observations(batch_size=8):

Model	Rouge-1-R	Rouge-1-P	Rouge-1-F	Rouge-2-R	Rouge-2-P	Rouge-2-F	Rouge-L-R	Rouge-L-P	Rouge-L-F	BLEU-1	BLEU-2	BLEU-3	BLEU-4
Deepseek-R1	0.6238	0.1955	0.2919	0.2375	0.0482	0.0783	0.4343	0.0675	0.1140	0.1011	0.0394	0.0167	0.0079
Deepseek-V3	0.5452	0.2670	0.3509	0.2031	0.0793	0.1106	0.3749	0.1191	0.1748	0.1965	0.0772	0.0373	0.0196
Qwen3-4B	0.4099	0.2749	0.3203	0.1651	0.0820	0.1056	0.3299	0.1359	0.1849	0.2271	0.0806	0.036	0.0196
Qwen3-0.6B	0.3840	0.2722	0.3095	0.1596	0.0835	0.1054	0.3302	0.1332	0.1825	0.2191	0.0824	0.0416	0.0251
Qwen2.5-VL-7B	0.4436	0.3177	0.3600	0.1756	0.1075	0.1283	0.3419	0.1774	0.2237	0.2903	0.1134	0.0605	0.0377
ThermalGuard_v1_4	0.5014	0.2967	0.3564	0.2096	0.0958	0.1233	0.3492	0.1378	0.1840	0.2469	0.0941	0.0462	0.0271
ThermalGuardV_v1_1	0.4503	0.3195	0.3591	0.1764	0.1104	0.1284	0.3186	0.1663	0.2061	0.2843	0.1056	0.0538	0.0328

Disclaimer

This model (ThermalGuardV) is a research-oriented AI tool independently developed for materials science applications. The model's outputs should be considered as informational suggestions rather than professional advice, and users are advised to verify critical materials science information through authoritative sources.