Truthseeker87/solarhive-community-solar-multimodal

SolarHive Community Solar Dataset

Canonical training corpus for the SolarHive family of fine-tuned Gemma 4 models. 1,727 rows (1,713 text + 14 image-grounded).

A combined text + sky-image training corpus for community solar energy intelligence. Built to fine-tune Gemma 4 into an AI energy advisor for residential solar microgrids — answering questions about production, storage, grid mix, weather impact, maintenance scheduling, and cross-source planning, with native function-calling for live data retrieval.

Total Examples	1,727 rows (1,713 text + 14 image-grounded)
Text-only Q&A	1,530 unique (algorithmic + hand-crafted, after deduplication)
Tool-calling Q&A	183 (across positive, negative, follow-up, and failure-recovery cases)
Image-grounded Q&A	14 turns from 7 manually-labeled sky photos
Locations	Ann Arbor, MI + San Mateo, CA
Data Sources	Open-Meteo, NREL PVWatts v8, OpenWeatherMap, EIA Open Data v2
Estimated Tokens	~347,789
License	MIT
Schema	Hugging Face Dataset with Image feature + messages (JSON chat string) + modality ("text" | "image")

---

Dataset Description

What's in it

The corpus is built from three complementary sources, all human-auditable. Composition (sums to 1,727; 5 duplicates removed during assembly):

Component	Count	Notes
Hand-crafted Q&A	413	15+ US cities, 9 energy domains (sky, storage, panel health, consumption optimization, community + grid coordination, multi-step reasoning, emergency + resilience, seasonal forecasting, tool-calling format)
API-grounded Q&A	~1,117	Six topical categories (atmospheric / production-vs-baseline / multi-source planning / weather context / battery & grid coordination / forecast horizon), joined on (location, hourly timestamp) from Open-Meteo, PVWatts, OWM, EIA
Tool-calling examples	183	Decomposed below
Image-grounded Q&A	14	7 manually-labeled Ann Arbor sky photographs
Total	1,727	1,713 text + 14 image

Tool-calling decomposition follows Ross et al. (2025) When2Call: When (not) to Call Tools, a taxonomy designed to address the 9–67% tool-hallucination rates that public tool-calling datasets exhibit because they lack follow-up and unable-to-answer examples:

• 106 (b) should-call cases (single-tool, multi-tool chains, full 5-tool community audits across get_weather / get_solar_production / get_battery_state / get_grid_status / get_nrel_pvwatts_baseline)
• 53 (a) should-not-call cases (general knowledge — teaches when NOT to call tools)
• 10 (d) unable-to-answer cases (questions outside available tools — model names limit + redirects)
• 6 (c) follow-up clarification cases (insufficient input to call a tool — model asks for the missing detail)
• 8 failure-recovery sequences (graceful handling when a tool returns an error)

Every numeric claim in an algorithmic answer traces back to a real API response — no LLM-generated numbers.

Inference-time validation. The companion inference notebook runs three held-out probes — one each in the (b), (c), (d) categories above — to verify that models fine-tuned on this corpus actually exhibit the expected behavior at runtime, not just in the training data. The 26B A4B family achieves 3/3 on these probes; the E4B family achieves 2/3 (passes (b) + (d), fails (c) by auto-filling location). Models trained without the unable-to-answer + follow-up clarification categories fail (c) + (d) by hallucinating tools or auto-filling defaults — exactly the failure modes the paper documents at 9–67% rates in untrained community models.

Tool-result message format used at training time (and required by the model at inference): a TWO-message sequence per tool round — {"role": "assistant", "tool_calls": [{"function": {"name": ..., "arguments": ...}}, ...]} followed by one {"role": "tool", "name": "<fn>", "content": "<JSON string>"} per result. This format is generated by the data-generation pipeline, validated by the fine-tune SFT preprocessing layer, and emitted at inference by both the cloud agentic loop (solarhive_inference.py) and the local Ollama HTTP raw-mode harness — inference matches the training distribution exactly. This differs from Google's vanilla function-calling docs (which use a single message with tool_calls + tool_responses keys); we use the 2-message form because that's what the model was fine-tuned on.

Image-grounded Q&A. Each of the seven Ann Arbor sky photographs is manually labeled with cloud type (clear / partly cloudy / cloudy / overcast) and percentage cloud cover (0–100), paired with ~2 question-answer turns each. Numeric claims (expected solar production from observed sky conditions) trace back to the cloud-cover label via the same temperature-derated GHI formula used throughout the dataset.

Honest aggregation in multi-step examples

Multi-step planning answers (weekly summaries, seasonal maintenance plans, annual reports) report aggregate production using two complementary statistics over daylight hours:

• Average (mean of daytime production) — composes correctly with cost and CO2 calculations
• Typical (median of daytime production) — robust to skew, represents what most days actually look like

Both are presented together (e.g., "Daytime efficiency: avg 27.5% (typical 24.2%) of 72 kW capacity"). Restricting to GHI > 0 hours avoids the misleading 24-hour mean that would otherwise halve the visible efficiency by including 12 hours of nighttime zeros.

Multimodal example format

{
  "image": "<Hugging Face Image feature — JPEG bytes, or null for text rows>",
  "messages": "[{\"role\":\"system\",\"content\":\"You are SolarHive...\"},{\"role\":\"user\",\"content\":[{\"type\":\"image\"},{\"type\":\"text\",\"text\":\"What does this sky tell you about solar production right now?\"}]},{\"role\":\"assistant\",\"content\":\"Cumulus at ~30% coverage. Expected production: ~52 kW from your 72 kW system. Battery should absorb the surplus during gaps.\"}]",
  "modality": "image"
}

The messages field is a JSON-serialized chat list. The modality column lets training filter to text-only or image-only subsets for ablation runs. Text-only rows use image: None, modality: "text" and follow the same system + user + assistant chat structure.

---

Data Sources

Live API grounding for text Q&A

Source	Data	Access	Coverage
Open-Meteo Archive	Hourly GHI (W/m²), DNI, DHI, low/mid/high cloud cover, temperature, wind, humidity	Free, no API key	~3 years per location, hourly resolution
NREL PVWatts v8	Monthly + hourly typical-meteorological-year production estimates	Free API key	Cross-validation per location
OpenWeatherMap	Real-time weather snapshot (temperature, clouds, wind, humidity, sunrise/sunset)	Free API key	Per location, snapshot at generation time
EIA Open Data v2	Grid fuel mix, renewable percentage, CO2 intensity (hourly)	Free API key	MISO + CAISO hourly, ~30-day rolling window

All four sources are joined on (location, hourly timestamp) so each multi-source example carries co-occurring grounding (same location, same hour). For dates outside the EIA Open Data API's recent-month window, grid stats fall back to the regional mean across the most recent EIA data — every row carries some grid value, with date-matched values used wherever the API window allows.

Image data — project archive

Source	Volume	Labels	Access	License
Project archive (Ann Arbor sky photographs)	7 unique photos	Manual labels: cloud type (clear / partly_cloudy / cloudy / overcast) + cloud_pct (0–100)	Local	Project-owned (MIT)

The image corpus is intentionally small. Earlier plans considered larger external corpora — the SWIM corpora from the National University of Singapore (rejected because they're CC BY-NC 4.0 and gated behind a registration form) and the NREL Solar Radiation Research Laboratory's legacy Sky Imager archive (rejected because it stopped publishing raw images in May 2017) — but none provided a license-compatible direct download path. The shipped image set is smaller than originally planned but every label is human-confirmed and every paired Q&A is grounded in the same temperature-derated GHI formula used elsewhere in the dataset.

Data principles

Full methodology in data_principles.md. Two core commitments:

1. No LLM-generated numbers. Every numeric claim traces to a real API response, a documented formula, or a manually-recorded label.
2. Cross-source coherence. Examples that quote multiple data sources do so at the same (location, timestamp) wherever the underlying APIs allow.

---

How to Use

from datasets import load_dataset

# Full dataset (text + image)
dataset = load_dataset("Truthseeker87/solarhive-community-solar-multimodal", split="train")

# Filter to text-only rows
text_only = dataset.filter(lambda x: x["modality"] == "text")

# Filter to image rows
image_only = dataset.filter(lambda x: x["modality"] == "image")

For supervised fine-tuning with Unsloth, see the training pipeline in solarhive_finetune.py. The shipped fine-tune is text-only: image rows are skipped at data-prep time and the pipeline pre-renders text rows via processor.apply_chat_template(...) for TRL's default text collator. VQA at inference time uses the base Gemma 4 model's pretrained vision encoder (~150M params for E4B, ~550M for 26B A4B per the official model card) — our LoRA targets only the language-model linear layers, matching the Vertex AI Gemma 4 SFT recipe documented in the Hugging Face blog, which explicitly freezes both vision and audio towers during text-focused fine-tuning. The dataset's image schema is preserved so a future multimodal fine-tune (post-hackathon) can re-enable image rows without changing the corpus.

---

Diagnostic Charts

The data-generation pipeline emits 14 diagnostic charts that validate data quality, reveal geographic patterns, and cross-validate between independent sources. All charts are produced from live API responses during data generation — no synthetic numbers. Highlights:

Chart 1: GHI distribution by location (Ann Arbor median 265 W/m² vs San Mateo 364 W/m²)	Chart 2: Average hourly production curve, mean ± 1 std
Chart 9: Open-Meteo vs NREL PVWatts monthly cross-validation	Chart 11: Average fuel mix — MISO vs CAISO grid composition
Chart 13: Irradiance decomposition — total GHI separated into direct-beam (DNI) and diffuse (DHI) on a clear summer day	Chart 14: Vertical cloud-cover composition by month — low (<3 km) / mid (3–8 km) / high (>8 km) stratification

The full 14-chart catalog (irradiance, hourly production, seasonal heatmap, temperature derating, feature correlation, cloud cover by season, seasonal production boxplots, GHI vs production scatter, monthly PVWatts cross-validation, OWM snapshot, EIA grid mix, renewable% + CO2 time-series, irradiance decomposition, vertical cloud stack) is in the GitHub repo at datagen_charts/.

---

Companion Models

Model	Repository	Notes
SolarHive 26B A4B LoRA	solarhive-26b-a4b-lora	Cloud-tier model, 8/8 on the held-out agentic-loop benchmark
SolarHive 26B A4B NF4	solarhive-26b-a4b-nf4	Pre-quantized cloud variant for resource-constrained Hugging Face Spaces
SolarHive E4B safetensors	solarhive-e4b-ollama	Edge-tier safetensors, source for Ollama deployment
SolarHive E4B GGUF	solarhive-e4b-gguf	Edge-tier Q4_K_M GGUF + multimodal projector, 10/10 on the same benchmark via Ollama on a 16 GB CPU laptop

The shipped fine-tune trains text-only on this corpus (image rows skipped at the data-prep layer). VQA at inference uses the base Gemma 4 model's pretrained vision encoder, which is unmodified by our LoRA — see the "How to Use" section above for citations.

---

Run Provenance

Field	Value
Generator	solarhive_datagen.py
Generated at	2026-04-30 01:54 UTC
Total examples	1,727 (1,713 text + 14 image)
Duplicates removed	5
Estimated tokens	~347,789

---

Links

• GitHub: youshen-lim/the-gemma4-good-hackathon-solarhive
• Kaggle: The Gemma 4 Good Hackathon

Built with Gemma 4 in Ann Arbor, Michigan. April 2026.

Gemma is a trademark of Google LLC.