aegis-geo-mind-qwen2.5-7b

Aegis-Geo-Mind (Qwen2.5-7B, QLoRA)

A Qwen2.5-7B-Instruct model fine-tuned with QLoRA on a curated geology and Earth-science corpus, with an emphasis on petroleum geology (sedimentology of petroliferous basins, reservoir stratigraphy, well logging, and seismic interpretation).

Intended use

Educational and research assistant for geology and petroleum-geology concepts. It produces fluent, domain-appropriate explanations using correct terminology across structural geology, stratigraphy, sedimentology, petroleum systems, and well-log interpretation.

Training

• Base model: Qwen/Qwen2.5-7B-Instruct
• Method: QLoRA (4-bit), LoRA rank 16
• Sequence length: 1024 tokens
• Corpus: filtered geoscience QA (public GeoSignal subset) combined with curated general-geology and petroleum-geology QA pairs, deduplicated and quality-filtered.

Limitations (please read)

This is a supervised fine-tune without retrieval augmentation. It reliably reproduces the style and structure of expert geological reasoning, but it can state specific facts, numbers, and ratios incorrectly with high confidence. Known weak areas include:

• Kerogen type classification (H/C and O/C ratios can be inverted)
• Numeric ranges (oil-window temperatures, burial depths, maturity cutoffs)
• Occasional invented specific values

For any factual or numeric use, verify against authoritative sources (e.g. Tissot & Welte, Selley, Schlumberger logging references). A retrieval-augmented (RAG) version addressing these gaps is planned.

Not for operational decisions

This model must not be used as a sole basis for exploration, drilling, or any operational geological decision.