TomatoPGFM

TomatoPGFM v0.1.0 release metadata. Verify the SHA-256 value after downloading the inference artifact before loading it.

TomatoPGFM is a graph-conditioned genomic foundation model pretrained on a 66-accession tomato pangenome. It contains 479,195,678 unique parameters, with approximately 148,625,438 parameters active per token under top-2 expert routing. The backbone comprises 10 bidirectional Mamba2 blocks and five local-attention mixture-of-experts blocks.

Model inputs

Sequence input uses reverse-complement-folded, non-overlapping 6-mer tokens. The graph interface accepts eight token-aligned channels in this order:

1. log-transformed graph-segment length;
2. GC fraction;
3. log-transformed in-degree;
4. log-transformed out-degree;
5. log-transformed accession carrier count;
6. gene-tag indicator;
7. repeat/TE-tag indicator;
8. is_random_only indicator.

Five channels varied in the production pretraining shards. The two annotation-indicator channels and is_random_only channel were constant zero in those shards. Optional within-window adjacency is supplied separately through edge_index.

Training data

The pretraining panel contained 66 tomato accessions and 54.65 Gb of assembled sequence. The assemblies originated from the public resources associated with NCBI BioProjects PRJNA733299 and PRJNA809001, CNCB BioProject PRJCA030093, NCBI BioProject PRJNA1201608 and Zenodo DOI 10.5281/zenodo.17878268. LA1974 and MicroTom were reserved for downstream evaluation and were not used for graph construction or pretraining. The third-party raw assemblies and annotations are referenced by accession and are not redistributed in this model repository.

Evaluation context

The accompanying manuscript reports graph-input sensitivity on windows sampled from the 66-accession training panel, chromosome-separated frozen probes, LoRA adaptation on LA1974 and MicroTom, pooling sensitivity and inference efficiency. The external-accession gene/CDS evaluations used the graph-pretrained TomatoPGFM checkpoint with external graph input disabled. A zero-feature GraphAdapter-on condition is an implementation-path control and is not graph-conditioned inference on those accessions.

Files

File	Purpose
model.safetensors	Tensor-only FP32 inference weights exported from the production model state
INFERENCE_WEIGHTS_MANIFEST.json	Source-checkpoint identity, tensor count, dtype counts and tied-weight metadata
model_final.yaml	Architecture configuration used by the public TomatoPGFM code
train_final.yaml	Five-stage optimization curriculum
pretraining_final.json	Public path-neutral pretraining snapshot
tokenizer_vocab.json	Production reverse-complement-folded 6-mer vocabulary
tokenizer_config.json	Tokenizer contract and special-token metadata
config.json	Repository-level model metadata
load_checkpoint.py	Minimal strict-loading example
EXPECTED_SHA256SUMS.tsv	Expected identity of the core model, tokenizer and configuration files
LICENSE	Apache-2.0 terms for the released inference-weight package

The repository is not Transformers-native and does not claim compatibility with AutoModel.from_pretrained. The supplied safetensors file was generated from the production checkpoint's model state; it is not a renamed PyTorch pickle.

Loading

Install the public TomatoPGFM code and a CUDA/PyTorch-compatible mamba-ssm build, then run:

python load_checkpoint.py \
  --checkpoint model.safetensors \
  --config model_final.yaml \
  --device cuda

Only load the official weights after verifying their SHA-256. load_checkpoint.py uses safetensors.torch.load_file and restores the tied mlm_head.weight -> embed.weight alias before strict state-dict loading.

Intended use

TomatoPGFM is released for research on tomato genomic representation learning, graph-conditioned masked-token prediction and downstream genomic sequence analysis. Outputs require task-specific validation before biological interpretation or breeding decisions.

Version and identity

• Version: v0.1.0
• Inference-weight filename: model.safetensors
• Bytes: 1,916,843,008
• SHA-256: 462f3bfc7178c4558fe993ef579d925da6a54a7448c975b8488bf74d0ac36f7c
• Tensors: 604
• Stored values: 479,195,678 FP32 values

Links and citation

• Source code: https://github.com/tomatoai2026/TomatoPGFM
• GitHub release: https://github.com/tomatoai2026/TomatoPGFM/releases/tag/v0.1.0
• GitHub software Version DOI: https://doi.org/10.5281/zenodo.22035724
• GitHub software Concept DOI: https://doi.org/10.5281/zenodo.22035723
• Model Version DOI: https://doi.org/10.5281/zenodo.22032734
• Model Concept DOI: https://doi.org/10.5281/zenodo.22032733
• Processed evaluation resources: https://doi.org/10.5281/zenodo.22033079
• ModelScope: https://modelscope.cn/models/turgun/TomatoPGFM
• Manuscript: bioRxiv preprint; add the article DOI after posting

Please cite the TomatoPGFM manuscript and the versioned model DOI. Model-weight license: Apache-2.0.

The optimizer-bearing training-resume checkpoint is intentionally not included in this public model release. The source-checkpoint hash in INFERENCE_WEIGHTS_MANIFEST.json records export provenance only.