biohack — osteosarcoma multi-omic imputation

Fine-tuning Muse-Glimmer-30B to fill a 20% random mask of Sid Sijbrandij's public osteosarcoma dataset (s3://sid-sijbrandij-osteosarc-dataset), plus classical baselines and an Esperanto ET-SoC-1 card experiment. Start with INVENTORY.md: every dataset, label set, baseline, fine-tune and hold-out experiment, with sizes, sha256, run status and where each file lives.

Model weights: every LoRA adapter (sid_full, stopped at 50%; the five lomo740_* runs; three smoke_* tests), with optimizer state and tokenizer, is on Hugging Face: huggingface.co/likelytobelaura/osteo-glimmer-lora.

What is in this repo

Code, reports and scores; every model answer file; the per-hold-out prompt and gold-label slices; and a full text copy of the Modal volume osteo-data in 11_modal_osteo_data/ (training set train/sid_full.jsonl, eval prompts prompts/baseline.jsonl, all eval shards). Not in git (too large or not needed; listed with sha256 in inventory.json): the full gold labels, raw bucket files, adapter weights (on HF), card result matrices (*.f32), raw ET-SoC-1 simulator logs and the public MNIST datasets used on the card.

folder	contents	in git
01_sid_full_dataset/	SFT training set (11,824 rows) + raw bucket files	file list only; training set is in 11_modal_osteo_data/train/
02_golden_labels/	gold values for all 24,634 masked targets	build script, README, provenance; labels local only
03_baseline_prompts/	the 15,964 eval prompts	no; same prompts are in 11_modal_osteo_data/prompts/
04_baseline_responses/	Muse-Glimmer-30B zero-shot answers	yes
05_baseline_report/	score.py report on the zero-shot baseline	yes
06_heldout_experiments/	leave-one-data-type-out runs: answers, per-type baseline slices, hold-out and off-target scores	yes (weights on HF)
07_finetune_sid_full/	sid_full LoRA (all data, stopped at step 739/1478): answers, scores, configs/logs, dashboard source	yes (weights on HF)
08_other_baselines/	Muse Spark 1.2 / 1.3 and NVIDIA NIM (no-context) answers and scores	yes (prompt files local)
09_classical_ml/	classical imputation sweep (kNN etc.), its predictions, and the osteo ET-SoC-1 card job	yes (*.f32 local)
10_esperanto_et_soc1/	everything from the AI Foundry lab account (aifoundry3): osteo card job, MNIST on the card, first tests	yes, except raw sim logs and datasets
11_modal_osteo_data/	full copy of the Modal volume osteo-data, same layout	all text; weights/optimizer/tokenizer on HF
12_dashboard/	dashboard build: build.py scores every run with the shared scorer and writes data/*.json; index.html is the page	yes
13_linear_probes/	MetaPhlAn log-abundance linear probe and microbiome-neuron hunt on Glimmer activations	yes (activations on Modal/HF)
14_crosstype_snapshots/	cross-type "Sid at timepoint X" snapshot fine-tune: scores, hint-copy analysis, scripts	yes (weights on HF xtype/)
scripts/	build, train, eval and scoring scripts	yes

Experiments

Every run is scored by score_compare.py on the same 24,634 targets of the 20% random mask (not a time split).

experiment	question	write-up / data
Glimmer zero-shot	how well does an untrained 30B model fill gaps from the entity's other values?	05_baseline_report/REPORT.md, 04_baseline_responses/
sid_full LoRA	does fine-tuning on the visible data help?	07_finetune_sid_full/
Leave-one-type-out LoRAs (lomo740_<type>, lomo740_all)	how much of the gain survives when a type is removed from training, and what does removing it do to the other types?	06_heldout_experiments/ (INVENTORY §6)
Context-free Glimmer	can the model predict a type with no observed values of it?	08_other_baselines/nim_noctx/
Muse Spark 1.3 / 1.2	does a larger hosted model know more, without fine-tuning?	08_other_baselines/muse_spark/
Classical imputation	how far do data-only methods get?	09_classical_ml/RESULTS.md, preds/
Cross-data-type prediction	which data types carry information about which others (shared genes, blood draws)?	09_classical_ml/RESULTS.md (cross section), cross_results.json, preds_cross/
Classical hold-out settings	can classical methods predict a type with none of it (S1), only other samples of it (S2), or only the sample's own values (S3)?	09_classical_ml/HOLDOUT.md, holdout_results.json, preds_holdout/, scripts/holdout.py
ET-SoC-1 card run	can the kNN similarity step run on the Esperanto card?	09_classical_ml/card/, 10_esperanto_et_soc1/

All scores side by side: 12_dashboard/data/ (types.json best method per type, leaderboard.json every run on every type, transfer.json the held-out types, offtarget.json hold-out effects with bootstrap CIs, classical_holdout*.json).

Keeping it current

• python3 inventory.py: refreshes the mirrors and rewrites INVENTORY.md / inventory.json (--offline skips Modal and the lab machine).
• git_sync.sh: commits and pushes the refresh (fast-forward only; refuses on credential patterns or files over 50 MB). watch_sweep.sh runs both every 10 minutes while experiments are running.
• ~/biohack/.venv/bin/python hf_upload.py / hf_volume_backup.py: upload adapters to HF, checking for truncated weights and verifying each file afterwards. vol_audit.py checks that every file on the Modal volume is on GitHub or HF.