code7monkey/cho-state-chronos2-final350-full

CHO state forecasting: full

Base revision: 29ec3766d36d6f73f0696f85560a422f50e8498c. Method: full. Seed: 42. Exported checkpoint step: 350. Training budget: 500; final deployment is Full step 350. LoRA weights are merged before export. No additional fitting for zero_shot. 24h forecast on 6h grid: VCD, viability, glucose, lactate, ammonium.

Source and license

Uhlendorff et al. (2026), Dataset Based on Chinese Hamster Ovary (CHO) Cultivations including Turbidity, Permittivity, O2 and CO2 Measurements, v1. https://doi.org/10.5281/zenodo.20829178 . Source data: CC-BY-4.0. Changes: causal right-edge binning; limited past-only fill for inline covariates. Original data is not bundled. Code/model and source-data licenses are separate.

Test results

method	scope	target	n	n_experiments	mae	rmse	persistence_mae	normalized_mae	coverage_80	mean_interval_width	zero_shot_mae	improvement_vs_zero_shot_pct
full	fed-batch	ammonium_mm	22	1	1.5485	1.7582	1.1627	1.3318	0.3636	2.7823	1.4943	-3.6275
full	fed-batch	glucose_g_l	22	1	0.7150	1.0134	0.5638	1.2682	0.3636	1.8207	0.8950	20.1092
full	fed-batch	lactate_g_l	22	1	0.9537	1.2401	0.9221	1.0343	0.5455	2.1864	1.4429	33.9062
full	fed-batch	vcd_cells_ml	22	1	877675.4773	1294423.8394	1293787.8788	0.6784	0.7273	2654968.2898	1840893.6894	52.3234
full	fed-batch	viability_pct	22	1	3.5968	4.7217	8.1898	0.4392	0.9091	13.3540	5.2619	31.6445

Limitations

Single-seed, small experiment-group holdout. Paired reactors stay in one split. No Keytruda production data and no validated transfer to 2000L. No titer, HCP, DNA or product-quality labels. No optimal harvest-day prediction. Deviation flags are not contamination diagnoses. Harvest review threshold is illustrative. Not validated for manufacturing control or batch release.

Reproduce and infer

See preprocessing.json, split.json, requirements.txt, environment-lock.txt and cho_model.py.