What are limits of language models for parsing handwritten notes and formulas?

[Kenneth_B]

I’ve been trying to use a large language model to help me parse and categorize decades of old, handwritten lab notes in my field, but I’m hitting a wall. The model keeps confidently misinterpreting chemical formulas and unit abbreviations from the specific notation my old professor used. I’m starting to wonder if this approach is fundamentally flawed for such a niche, domain-specific task without a massive and tailored training set. Has anyone else run into this problem with specialized scientific literature?

[LukeUL]

I’ve tried a similar setup and the confidence with which it locks onto a wrong chemical formula is maddening. Handwriting quirks, subscripts, and old abbreviations would turn into totally different compounds in a single pass. It didn’t feel like occasional slips; it felt baked in after a few pages. A handful of pages with metal salts and solvent codes kept tripping it.

[OliverT]

I wonder if the real bottleneck isn’t the model at all but the handwriting and the context. If a note says CuSO4 with a tiny dot that means something else in that lab, the tool can’t guess. OCR misreads happen, the symbol dictionary is missing, you end up with a confident but wrong read. I found it cheaper to run a human-in-the-loop workflow first and only then think about automation.

[EthanA]

I did a tiny pilot: about 300 lines labeled, and I found around 60 entries that looked readable, but about 15 were wrong because of stray strokes or ligatures. I built a small glossary of the professor’s shorthand and added a few post-processing checks: if a parsed formula doesn’t match known patterns, tag it for review. The gains were real but not trustworthy for a large archive yet.

[Jacob_M]

I kept thinking maybe the issue isn’t the parsing at all but the dataset itself. The notes aren’t a clean corpus—experiments drift, units swap around, and context is scattered. I tried focusing on a narrow slice first, like standard compounds and units, and that helped a bit, but the moment you switch labs or eras everything changes. Is the problem really the niche notation, or is the tool just not built for this kind of uneven data?