lylybig8/routing-analysis-anvil-six-models

Routing-guided language adaptation artifacts

This repository is a path-preserving research snapshot for Hausa and Yoruba adaptation experiments based on AIDC-AI/Marco-Nano-Base. It contains model weights, resumable checkpoints, training and evaluation code, logs, and raw LightEval results collected on the Anvil cluster.

The directory hierarchy follows the original routing_analysis project so that scripts and recorded paths remain easy to trace. No symbolic links are uploaded: every Hub entry is a regular file.

Complete models

• finetuning/phase2_full/checkpoints/marco_nano_base/hau_Latn_100k_full
• finetuning/phase2_full/checkpoints/marco_nano_base/yor_Latn_50k_full
• finetuning/phase2_full/checkpoints/marco_nano_base/yor_Latn_50k_exp3_train_only_yor_pretraining_corpus_top8_experts

These directories are preserved in full, including retained intermediate checkpoints, optimizer state, trainer state, tokenizer files, and completion or verification metadata.

Resumable partial runs

• Hausa 500K Top-8: checkpoint-1682 of 7813
• Yoruba 100K Full-FT: checkpoint-519 of 1563
• Yoruba 50K rank-64 Top-8 LoRA: checkpoint-335 of 782

Only the selected latest recoverable checkpoint is included for each partial run, together with the root-level run configuration, audit, pause, and resume metadata required to understand the checkpoint.

Evaluation artifacts

The repository includes model-specific scripts and results for:

• Belebele: fixed 34-language retention panel plus Hausa or Yoruba target task.
• GlobalMMLU: fixed 23-language retention panel plus Hausa or Yoruba target task.
• LightEval v0.10.0 task definitions and the shared resumable runner.

Raw result JSON, details Parquet files, logs, validation records, and completion markers are kept under their original reproduce/ paths.

Training inputs and provenance

The exact training splits, manifests, and the shared Marco-Nano base snapshot used by the runs are included under their original paths. Dataset sources:

• FineWeb-2: HuggingFaceFW/fineweb-2, ODC-By 1.0, also subject to Common Crawl terms.
• Wura: castorini/wura, Apache-2.0.
• Marco-Nano-Base: Apache-2.0 according to its model card.

See _artifact_manifest/ for the source mapping, exact byte counts, file inventory, copied-file checksums, and validation report generated before the Hub upload.

Size

The snapshot is approximately 1.03 TiB. Most of the size comes from preserved optimizer states and intermediate checkpoints, which are intentionally retained to satisfy the archival and resumption requirements of these experiments.

Intended use

This is a research artifact rather than a single flattened inference model. Choose an individual completed model directory for inference. Partial checkpoints should be resumed with their matching scripts, base model, data, and environment recorded in this repository.