Phytomni-SFT

Supervised fine-tuning corpus for the Phytomni multi-agent system. The corpus is used to fine-tune both Phytomni LLMs — Phyto-Chatbot and Phyto-Reasoner — after continual pre-training on Phytomni-Pretrain.

Construction

The release contains 20,982 examples, split evenly between the two models: 10,491 for Phyto-Reasoner and 10,491 for Phyto-Chatbot. Each per-model half mixes plant-science domain examples with general-domain examples (math, programming, reasoning, medicine, social sciences, and general knowledge). The 14-way scene field distinguishes (a) which model the example targets and (b) which task bucket it covers.

Task bucket	Reasoner rows	Chatbot rows
knowledge-article	1,937	1,937
knowledge-review	1,815	1,815
knowledge-trace	1,648	1,648
knowledge-gene	1,500	1,500
knowledge-id	1,333	1,333
analysis	1,274	1,274
data	984	984
Total	10,491	10,491

Schema

Field	Type	Description
messages	list of {role: string, content: string, calculate_loss: int64}	A multi-turn chat sequence. calculate_loss == 1 marks turns whose tokens contribute to the SFT loss; calculate_loss == 0 marks turns that should be masked out (typically prompts and tool responses).
scene	string	One of 14 values: <model>-<bucket> where <model> ∈ {reasoner, chatbot} and <bucket> is one of the 7 task buckets above. Use it to filter the corpus per target model and per skill.

How to use

from datasets import load_dataset

ds = load_dataset("Phytomni/Phytomni-SFT", split="train")

# Filter to a single model
reasoner = ds.filter(lambda r: r["scene"].startswith("reasoner-"))
chatbot  = ds.filter(lambda r: r["scene"].startswith("chatbot-"))

# Apply your chat template, then mask losses where calculate_loss == 0
sample = ds[0]
for turn in sample["messages"]:
    print(turn["role"], "(loss=", turn["calculate_loss"], "):", turn["content"][:120])

When constructing training labels, set the loss mask to zero on every token whose source turn has calculate_loss == 0, and to one elsewhere. This convention preserves the original prompt/response boundaries authored during corpus curation.

Citation

@article{phytomni2026,
  title  = {Phytomni: An agentic AI accelerating plant research from discovery to design},
  author = {Phytomni Team},
  year   = {2026},
  note   = {Manuscript in preparation; citation TBD until publication.}
}

Links

• Project: https://github.com/Phytomni/Phytomni
• Agent code: https://github.com/Phytomni/Phytomni-Bot

License: GPL-3.0

Version

v0.1.0