---
license: apache-2.0
base_model: mistralai/Magistral-Small-2506
language:
  - en
  - zh
library_name: peft
tags:
  - chemistry
  - cheminformatics
  - information-extraction
  - lora
  - not-yet-released
pipeline_tag: text-generation
---

# formulation-record-extractor (placeholder)

A LoRA adapter (to be released) that maps messy chemistry experiment records into a
structured formulation schema, and supports a tool-using "chemical reasoning"
Copilot. By [Formulith](https://www.formulith.com).

> 本專案部分研發由**數位發展部數位產業署 115 年 AI 算力平台**支持。
> Part of this work is supported by the **AI Computing Platform (2026), Administration for Digital Industries, Ministry of Digital Affairs, Taiwan**.


## Status

**Weights not yet uploaded.** The adapter will be released after the grant period.
Until then this card documents intended scope.

## Training data

Public chemistry data + the **synthetic** `Formulith/formulation-records-instruct`
set only. **No customer data and no proprietary formulations are used to train any
released model.** Customer-specific tuning, if any, stays private and is never
published.

## Intended use / limitations

- Use: structure lab records, normalize names/units, draft explanations of
  optimizer (BO/XAI) outputs for chemists.
- The model is paired with **RDKit validation** (fingerprints, stoichiometry /
  atom-conservation) — it must not be relied on for un-validated chemistry.
- Base model: Magistral Small 2506 (Apache-2.0), open and non-PRC.

## License

Apache-2.0 (adapter). Base model under its own license.