This is a LoRA adapter for unsloth/medgemma-4b-it-bnb-4bit. It was fine-tuned using text-only English-language Opentrons OT-2 protocol descriptions to produce machine-readable JSON containing the protocol title, categories, labware, pipettes, modules, and reagents. It extracts protocol metadata; it does not generate executable robot code.

This model is a fine-tuned version of unsloth/medgemma-4b-it-bnb-4bit. It has been trained using TRL.

Model documentation: MedGemma

Base model: unsloth/medgemma-4b-it-bnb-4bit

Github repo: yahyamomtaz/Meddemma-opentrons-fine-tuning

Data Source: Opentrons Protocol Library.

Research and assistive use only. This adapter does not generate, validate, or approve executable robot protocols. Every output must be reviewed by a qualified laboratory professional and validated in the Opentrons Protocol Designer / simulation environment before use.

Intended use

Use this model as a draft metadata-extraction assistant for natural-language protocol descriptions, for example inventory planning, protocol-library search, or human-in-the-loop annotation.

It is not intended for clinical decision-making, diagnostic use, autonomous liquid handling, safety-critical lab operations, or use where an incorrect labware, reagent, module, or pipette selection could cause harm.

Task and output schema

Input: a natural-language Opentrons protocol description.

Output: one JSON object, without explanatory text or Markdown fences:

json

{
  "title": "string",
  "categories": ["string"],
  "labware": ["string"],
  "pipettes": ["string"],
  "modules": ["string"],
  "reagents": ["string"]
}

The strings are model predictions, not a validated controlled vocabulary. They may be generic, duplicated, omitted, or unsupported by the input.

Training configuration

Setting	Value
Base model	`unsloth/medgemma-4b-it-bnb-4bit`
LoRA rank / alpha / dropout	16 / 32 / 0.05
Epochs	3
Optimizer	Paged AdamW 8-bit
Learning rate	1e-4
Batch size / gradient accumulation	2 / 4
Maximum sequence length	2048
Hardware	NVIDIA H100

The final checkpoint at step 246 was also the selected best checkpoint, with eval_loss=0.35. The completed run reported train_loss=0.40. These loss values are optimization diagnostics, not task-quality or safety metrics.

Evaluation

Labelled evaluation split

The model was evaluated on all 164 source-disjoint records in data/eval.jsonl.

Field	Precision	Recall	F1
Categories	0.79	0.78	0.79
Labware	0.35	0.32	0.34
Pipettes	0.65	0.70	0.67
Modules	0.67	0.80	0.73
Reagents	0.42	0.18	0.25

structured-output

bash

{
  "title": "GNA Octea Prep",
  "categories": [
    "Sample Prep"
  ],
  "labware": [
    "Opentrons 96 Tip Rack 300 µL",
    "Opentrons 96 Tip Rack 1000 µL",
    "Opentrons 24 Tube Rack with Eppendorf 1.5 mL Safe-Lock Snapcap",
    "Opentrons 96 Filter Tip Rack 20 µL",
    "Opentrons 96 Well Aluminum Block with Generic PCR Strip 200 µL",
    "Opentrons 24 Well Aluminum Block with NEST 1.5 mL Snapcap",
    "Opentrons 6 Tube Rack with Falcon 50 mL Conical"
  ],
  "pipettes": [
    "P300 8-Channel Pipette (GEN2)",
    "P300 Single Channel Pipette (GEN2)"
  ],
  "modules": [
    "Magnetic Module",
    "Magnetic Module GEN2",
    "Temperature Module",
    "Temperature Module GEN2"
  ],
  "reagents": []
}

License and attribution

This adapter is a derivative of MedGemma. Its use and redistribution are governed by the applicable Health AI Developer Foundations terms. Do not replace this notice with a permissive license unless you have confirmed that doing so is compatible with the base-model terms and the data provenance.

The underlying data derive from Opentrons protocol-library and catalog artifacts. Verify the exact upstream revision and redistribution rights before publishing the dataset or its enriched fields.

Citation

The model has been trained using TRL.

bibtex

@misc{vonwerra2022trl,
	title        = {{TRL: Transformer Reinforcement Learning}},
	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
	year         = 2020,
	journal      = {GitHub repository},
	publisher    = {GitHub},
	howpublished = {\url{https://github.com/huggingface/trl}}
}

yayamomt/medgemma-4b-it-sft-lora-opentrons

MedGemma Fine-Tuning for Opentrons Protocol Metadata