Datasets:

jonpark0
/

contilab-cpu-study

license: mit
pretty_name: 'contilab CPU study: continual learning results'
tags:
  - continual-learning
  - lifelong-learning
  - catastrophic-forgetting
  - class-incremental-learning
  - benchmark
  - model-merging
  - knowledge-distillation
size_categories:
  - n<1K
configs:
  - config_name: cpu_main
    data_files: cpu_main.csv
    default: true
  - config_name: cpu_tuned
    data_files: cpu_tuned.csv
  - config_name: cpu_student
    data_files: cpu_student.csv

contilab CPU study: continual learning results

Per-run results of an empirical comparison of 12 continual-learning methods on 6 benchmarks, produced with contilab 0.3.1, a PyTorch framework for continual learning research. One row per run, with every scalar metric: final accuracy, forgetting, backward and forward transfer, the stability gap, task-recency bias, calibration, representation metrics, memory and wall time.

Results page (rankings, significance tests, findings) · Docs · GitHub · DOI 10.5281/zenodo.23906803

Files

File	Config	Rows	Contents	Experiment config
`cpu_main.csv`	`cpu_main`	324	Main study: every method with default hyperparameters on 6 benchmarks	`cpu_main.yaml` + `cpu_main_seeds.yaml`
`cpu_tuned.csv`	`cpu_tuned`	456	Tuned comparison: per-method grids on the two MNIST benchmarks, variant chosen on validation accuracy (`selected`)	`cpu_tuned.yaml`
`cpu_student.csv`	`cpu_student`	12	Teacher → smaller student: distillation into a 2×100 MLP	`cpu_student.yaml`

The main study has 324 runs: 312 done, 12 diverged, 0 failed. Seeds per benchmark: Split MNIST (Class-IL) 6, Split MNIST (Task-IL) 6, Permuted MNIST (Domain-IL) 6, Split CIFAR-10 (Class-IL) 3, Split CIFAR-100 (Class-IL) 3, text5 + Pythia-70M (Class-IL) 3.

from datasets import load_dataset

main = load_dataset("jonpark0/contilab-cpu-study", "cpu_main", split="train")

Columns

Identifiers: run_id, preset (benchmark and scenario), method, variant (swept hyperparameters), seed, status (done, diverged or failed). cpu_tuned adds selected (the variant chosen on validation accuracy) and cell (the comparison-axis part of the variant).
Metrics: every other column. Definitions and references are on the Metrics page. An empty cell means the metric is undefined for that run, for example forward transfer in Class-IL, stream metrics for consolidation methods, or anything for a diverged run.

Final accuracy (cpu_main, mean over seeds)

Method	Split MNIST (Class-IL)	Split MNIST (Task-IL)	Permuted MNIST (Domain-IL)	Split CIFAR-10 (Class-IL)	Split CIFAR-100 (Class-IL)	text5 + Pythia-70M (Class-IL)
Naive	0.196	0.980	0.805	0.181	0.049	0.097
Joint (upper bound)	0.950	0.993	0.909	0.398	0.118	0.529
EWC	0.197	0.983	diverged	diverged	diverged	0.105
LwF	0.228	0.990	0.817	0.173	0.047	0.105
Experience Replay	0.858	0.983	0.869	0.302	0.059	0.329
DER++	0.871	0.991	0.902	0.256	0.059	0.397
A-GEM	0.335	0.984	0.864	0.181	0.048	0.100
LoRA-per-task	0.184	0.986	0.619	0.137	0.015	0.024
Task Arithmetic	0.531	0.979	0.570	0.158	0.011	0.034
TIES-Merging	0.561	0.983	0.677	0.166	0.013	0.033
Multi-teacher distillation	0.863	0.986	0.773	0.262	0.040	0.452
Continual distillation	0.586	0.980	0.818	0.210	0.030	0.331

"diverged" marks a benchmark where every seed of the method diverged (its loss stopped being finite); such runs are kept with status = diverged.

How the data were produced

Hardware: 4 CPU cores (Intel Xeon @ 2.1 GHz), no GPU.
Vision: MLP 2×400 on the MNIST benchmarks (1 epoch per task), reduced ResNet-18 on Split CIFAR-10/100 (2 epochs). Language: Pythia-70M on text5 (AG News → Yelp → Amazon → DBpedia → Yahoo Answers), 1,000 training texts per task.
Replay and distillation methods keep 500 samples (vision) or 200 (text).
Reproduce with contilab run configs/experiments/<config>.yaml; the setup and every deviation from the design are listed on the Results page.

Citation

@software{contilab2026,
  author  = {Park, Yeochoon},
  title   = {contilab: a performance-first framework for continual learning benchmarking, distillation and model merging},
  year    = {2026},
  version = {0.3.1},
  doi     = {10.5281/zenodo.23906803},
  url     = {https://github.com/JonPark0/contilab}
}

License

MIT, like contilab. The benchmarks' source datasets keep their own licenses.