EvoAge — Data, Models and Run Artifacts

This repository hosts the large files that back the EvoAge project — raw source data, trained KGE models, the Neo4j graph dump, the fine-tuned LLM, and hypothesis-run artifacts. Anything too big for git lives here; the code lives on GitHub.

Documentation: the-ahuja-lab.github.io/EvoAge

---

Which folder do I need?

I want to…	Go to
Run the EvoAge chatbot / backend locally	DGL-EvoKG/
Rebuild the Knowledge Graph from scratch	kg_formation/data_collection/
Train the KGE models myself	kg_formation/EvoAge_training_files/
Skip training and use our trained embeddings	kg_formation/KGE_model_files/
Just query the finished KG in Neo4j	kg_formation/neo4j/
Reproduce Fig. 2f	kg_formation/shuffled_test_set/
Use the aging fine-tuned BioMistral	LLM_finetunning/
Inspect the hypotheses EvoAge generated	hypothesis_runs/

---

DGL-EvoKG/ — run the system locally

These are the model artifacts the backend loads at startup. Use this folder when you have cloned the GitHub repository and want to run and test the chatbot system locally.

After cloning, download these files into Backend/DGL-EvoKG/ in your clone — the Dummy_Input/ and HypothesisTesting/ sub-folders already ship with the repository, the Model/ and Node_Mapping/ folders do not.

pip install -U "huggingface_hub[cli]"

# run from the Backend/ directory of your clone
hf download gauravahuja77/EvoAge \
  --repo-type dataset \
  --include "DGL-EvoKG/Model/*" "DGL-EvoKG/Node_Mapping/*" \
  --local-dir .

Path	Size	Contents
Model/RESCAL_Evoage_121_12M_0/	11.7 GB	Served RESCAL model — entity/relation .npy embeddings + config.json
Model/entities_final_Ntype_updated.dict, relation_final*.dict	1.4 GB	Entity and relation ID dictionaries
Node_Mapping/node_id_mapping_EvoAge_121_12M.pkl	1.4 GB	Node ID → name/type mapping used to decode predictions
Dummy_Input/	small	head.list / rel.list — smoke-test input for a first link-prediction call
HypothesisTesting/Model/EvoAge_121_12M_to_many_KG.pt	29.7 GB	Full 1-to-many KG tensor, needed only for hypothesis testing
HypothesisTesting/Cutoff/	small	Score percentiles and per-relation cutoffs used to call a triple plausible
HypothesisTesting/InputHypothesis/	small	Example head_/rel_/tail_ triple lists for a hypothesis run
HypothesisTesting/JSONResult/hypothesis_run_stats.csv	small	Summary stats across hypothesis runs

The 29.7 GB .pt tensor is only required for the hypothesis-testing path. Plain link prediction needs Model/ and Node_Mapping/ only.

Point ROOT_DIR_PATH in your backend .env at the directory containing Model/, Node_Mapping/ and Dummy_Input/, and make sure MODEL_PATH, ENT_DICT_PATH, REL_DICT_PATH and NODE_MAPPINGS_PATH resolve to the files you just downloaded.

---

kg_formation/ — reproduce the KG and the training format

Everything needed to rebuild the EvoAge Knowledge Graph and re-run KGE training. The code that consumes these files is in the pipeline/ folder of the GitHub repository.

kg_formation/data_collection/ — raw source data

Start here. 51 archives (~88 GB) holding all the raw data downloaded and used in EvoAge KG formation — the 13 aging-focused and 37 general biomedical resources, plus databases_for_mapping.tar.gz for identifier reconciliation.

Sources include: agingatlas, agingbank, ageanno, ageannomo, agextend, cellage, genage, gendr, drugage, digitalagingatlas, metaboage, hald, evolf, CKG/ckg, harmonizome, hetionet, primekg, monarchkg, drkg, pharmkg, tarkg, ibkh, crossbar, pheknowlator, biograkn, biogrid, biosnap, string, stitch, bindingdb, chembl, drugbank, ttd, dtinet, mousenet, yeastnet, sgd, flybase, wormbase, zfin, mgi_do, hmd_hp, mirTARbase, mirnaatlas, mirnatissueatlas, bock, eSLDB, fic, immpat, phytohub, sms, widb, other_sources.

To reproduce the whole KG, run pipeline stages 01_data_collection through 06_tensor_building against this data:

01_data_collection → 02_data_processing → 03_relation_merge → 04_orthology_mapping
→ 05_kg_construction → 06_tensor_building

kg_formation/EvoAge_training_files/ — KGE training splits

For pipeline stage 07_model_training you need the training files directly — they are here (~46 GB), as four archives plus a tree.txt describing the layout:

Archive	Contents
building_evoage_kg_new.tar.gz	EvoAge KG train/valid/test splits + .pt tensors (1-to-1, 1-to-many, 121_12M)
building_aging_kg_new.tar.gz	Aging-only KG splits and tensors
building_biomedical_kg_new.tar.gz	Biomedical-only KG splits and tensors
mappings.tar.gz	entities_final.dict, relation_final.dict

kg_formation/KGE_model_files/ — trained KGE models

Training all models one-by-one on a single GPU takes 3–4 days. To skip that, use the already-trained embedding files and configs here (~65 GB). Each archive expands to <KG>/model_output/<Algorithm>/<RUN>/ with *_entity.npy, *_relation.npy and config.json.

Archive	Models inside
EvoAge_121_12M/model_output/Transe.tar.gz	TransE on the full EvoAge KG
Aging_121_12M.tar.gz, Aging_1_to_1.tar.gz	RESCAL on the aging-only KG
Biomedical_121_12M.tar.gz, Biomedical_1_to_1.tar.gz	RESCAL on the biomedical-only KG
EvoAge_1_to_1.tar.gz	RESCAL on the 1-to-1 EvoAge KG
1_per_species_test_EvoAge_121_12M.tar.gz	Per-species held-out test variant

tree.txt lists the full expanded layout, including the ComplEx / DistMult / RESCAL / RotatE / SimplE / TransE runs on EvoAge_121_12M.

kg_formation/shuffled_test_set/ — randomized test sets

Thirty randomized shuffled test files (shuffled_test_000_seed0.txt … shuffled_test_029_seed29.txt, ~2.2 GB each) plus the matching EvoAge_121_12M_KG_train_90.txt and EvoAge_121_12M_KG_valid_10.txt. Use these to reproduce Fig. 2f.

kg_formation/neo4j/ — Neo4j database dump

neo4j.dump.tar.gz (~30 GB) — a dump of the complete EvoAge KG. Use it if you want to query the finished graph without rebuilding it.

sudo systemctl stop neo4j
sudo cp neo4j.dump /var/lib/neo4j/import/
sudo neo4j-admin database load neo4j \
  --from-path=/var/lib/neo4j/import/ --overwrite-destination=true
sudo systemctl start neo4j

---

LLM_finetunning/ — Aging-KG fine-tuned BioMistral

All files and scripts for the BioMistral model fine-tuned on the Aging KG. Corresponds to pipeline stage 10_llm_fintunning.

Path	Size	Contents
Biomistral_Finetuned_model/BioMistral-Finetuned_model2/	14.5 GB	Fine-tuned weights — 3-shard safetensors, config.json, tokenizer, chat template, and an Ollama Modelfile
Biomistral_Finetuning_traning_files/Aging_121_12M_forfinetune.csv	0.4 GB	The instruction-tuning set derived from the Aging KG

---

hypothesis_runs/ — hypothesis-engine artifacts

Artifacts produced by the EvoAge multi-agent hypothesis engine. Each run is one hash-named directory containing three files:

• response.json — the agent's verdict and reasoning
• evidence_records.json — KG-derived evidence supporting or contradicting it
• triples.csv — the triples retrieved for that hypothesis

Path	Runs	Contents
right_inverse_hypo/right/runs/	101	The 100 biologically correct hypotheses
right_inverse_hypo/inverse/runs/	101	Their deliberately inverted counterparts — the negative control
yeast_pulsing/all_yeast_pulsing_hypothesis_runs/	953	Yeast pulsing case-study runs

---

Downloading

Whole repository (~419 GB — you almost certainly want a subset instead):

git clone git@hf.co:datasets/gauravahuja77/EvoAge

A single folder:

pip install -U "huggingface_hub[cli]"

hf download gauravahuja77/EvoAge \
  --repo-type dataset \
  --include "kg_formation/data_collection/*" \
  --local-dir ./EvoAge

Individual files can also be downloaded from the Files and versions tab in the browser.

---

License

MIT