SolarGPT-QA: Solar Education LLM

• SolarGPT-QA is a specialized large language model for heliophysics education, integrating LoRA-based domain adaptation and pedagogical fine-tuning to deliver precise and accessible answers.
• It employs domain-adaptive pretraining on thousands of peer-reviewed heliophysical articles followed by synthetic QA fine-tuning to enhance scientific accuracy and clarity in educational settings.
• Evaluations show a 75% human win-rate and improved student comprehension, demonstrating its effective balance between technical rigor and age-appropriate explanation.

SolarGPT-QA is a specialized LLM architecture for educational question answering in space weather and heliophysics, employing domain-adaptive pretraining and pedagogical fine-tuning on a LLaMA-3 base. The model targets the systematic challenge of providing both scientific rigor and age-appropriate explanations in domains such as solar flares, coronal mass ejections (CMEs), and space weather forecasting, where general-purpose LLMs have demonstrated substantial limitations in both technical accuracy and clarity for non-expert audiences. Its construction utilizes large-scale scientific corpora, synthetic and curated question–answer data, and a two-stage adaptation pipeline involving Low-Rank Adapter (LoRA) injection and supervised refinement with human-aligned style via Grok-3. SolarGPT-QA demonstrates superior performance to generic LLMs in zero-shot educational query settings and achieves competitive explanations compared to instruction-tuned baselines, suggesting a principled path toward a broader SolarGPT framework for science education and operational explainability (Chapagain et al., 17 Jan 2026).

1. Model Architecture and Domain-Adaptive Pretraining

SolarGPT-QA is implemented as a parameter-efficient domain adaptation of the LLaMA-3-8B model, consisting of a causal decoder-only Transformer (32 layers, hidden size ≈ 4 096, 32 attention heads, 4 096-token context). Instead of full-model finetuning, SolarGPT-QA introduces domain-specific knowledge via Low-Rank Adapters (LoRA) integrated on query and value projections within self-attention blocks. The LoRA adaptation is parametrized by rank $r=16$, scaling factor $\alpha=32$, and dropout $p=0.05$, with only adapter weights ($A \in \mathbb{R}^{r \times k}$, $B \in \mathbb{R}^{d \times r}$) updated through AdamW.

The domain-adaptive pretraining objective minimizes the causal negative log-likelihood

$$\mathcal L_{\mathrm{DAPT}}(\theta) = \mathbb E_{x \sim \mathcal T}\Bigl[\sum_{t=1}^T -\log p_\theta(x_t \mid x_{<t})\Bigr]$$

where the corpus $\mathcal{T}$ comprises 5 000–7 000 peer-reviewed articles from heliophysics and solar science journals (ApJ, ApJS, A&A, Space Weather, Frontiers in Astronomy & Space Sciences, etc.), preprocessed to remove citations, references, figure captions, and section headers, followed by tokenization.

2. Synthetic Data Generation and Pedagogical Fine-Tuning

Post-DAPT, SolarGPT-QA undergoes a second refinement stage using synthetic and human-aligned question–answer data. An initial set of ~20 000 QA pairs is generated from the cleaned corpus using GPT-4 prompts, requesting concept-checking questions and K–12-level answers in heliophysical domains. Automated filtering removes responses containing advanced mathematics, ungrounded claims, or >150 words, and enforces required topical coverage, yielding ~10 000 QA pairs.

Pedagogical fine-tuning is carried out with Grok-3, focusing on “storytelling” answers featuring analogies and clear logical signposting (“first..., then..., so...”). LoRA adapters are used as in DAPT, trained with AdamW (learning rate $2\times10^{-5}$, batch 16, 3 epochs). Curriculum strategies stratify QA data by difficulty, cycling from Easy to Hard to stabilize convergence. The supervised loss is the conditional negative log-likelihood:

$$\mathcal L_{\mathrm{QA}}(\theta) = \mathbb E_{(q, a) \sim \mathcal D_{\mathrm{QA}}}[ -\log p_\theta(a \mid q) ]$$

3. Evaluation Methodology and Performance Results

SolarGPT-QA is evaluated in zero-shot and few-shot settings on document-held-out, stratified QA pairs (Easy/Medium/Hard levels). Benchmarks include ChatGPT (10-shot instruction-tuned), ChatGPT (zero-shot), Grok (zero-shot), and other LLMs (Claude, Gemini, DeepSeek). Metrics include: human pairwise win-rate, student comprehension on a 15-person pilot (self-reported “understood”), and statistical significance (McNemar’s test).

Human win-rate comparison:

Model	Win Rate (%)
ChatGPT (instr-tuned, 10-shot)	82
SolarGPT-QA	75
ChatGPT (zero-shot)	60
Grok (zero-shot)	55
Claude	45
Gemini	42
DeepSeek	40

Student comprehension (of 15):

Model	# Students Understood
ChatGPT (instr-tuned, 10-shot)	15
SolarGPT-QA	12
ChatGPT (zero-shot)	9
Grok	9
Claude	7
Gemini	6
DeepSeek	6

Ablations (win-rate, comprehension):

Variant	Win Rate (%)	Comprehension (%)
Base LLaMA-3-8B	28	20 (3/15)
LLaMA-3 + DAPT	54	40 (6/15)
LLaMA-3 + QA fine-tuning	48	67 (10/15)
SolarGPT-QA (full)	75	80 (12/15)

Both DAPT and pedagogical fine-tuning individually contribute (Δwin^DAPT = 21%, Δwin^QA = 27%), but best performance is achieved with both (Chapagain et al., 17 Jan 2026).

4. Component Roles and Architectural Insights

• LoRA adapters: Allow parameter-efficient domain specialization without loss of general linguistic competence.
• DAPT: Injects specialized heliophysical lexicon, causal patterns, and phraseology, improving scientific grounding and terminology coverage.
• Grok-3 pedagogical fine-tuning: Imposes K–12-aligned explanatory style, narrative flow, and analogical reasoning, directly optimizing for educational clarity.
• Curriculum stratification: Smooths optimization across task difficulty, stabilizing learning and combating catastrophic forgetting.

Human and student studies indicate DAPT alone yields technically detailed but less accessible answers, while QA fine-tuning without DAPT lacks scientific depth. Their combination is required to achieve high win-rate and comprehension.

5. Limitations and Future Directions

Limitations of SolarGPT-QA include the relatively small pilot student study (n = 15, limited grade coverage), restriction to single-modal text (i.e., no image or time-series augmentation), and the absence of large-scale automatic scoring (BLEU, ROUGE, F1). Current generation does not utilize retrieval-augmented generation or adaptive curriculum targeting individual student profiles.

Future work directions include:

1. Scaling to larger LLaMA-3 models with expanded corpus for deeper scientific reasoning.
2. SolarGPT-Core, targeting expert-facing detailed querying.
3. SolarGPT-PredEx, integrating multimodal (imagery, time-series) fusion for joint prediction and explanation.
4. Retrieval-augmented systems and adaptive curricula for personalized instruction.
5. Larger-scale, diverse classroom validation and systematic metric analysis.

6. Significance and Broader Context

SolarGPT-QA represents an initial step in developing robust LLM-based educational assistants for domain-critical scientific fields where traditional LLMs falter in both accuracy and pedagogy. It demonstrates that parameter-efficient adaptation (LoRA), coupled with synthetic QA generation and human-aligned fine-tuning, can yield models that outperform standard zero-shot LLMs in expert-evaluable scientific correctness and K–12 accessibility without dependency on costly large-scale instruction-tuning or in-context demonstration (Chapagain et al., 17 Jan 2026). The SolarGPT design pattern is extensible to further sub-fields requiring explanation, scoring, and classroom adaptability.