Model names

Released models are named

strands-decider-{size}-{base}-v{N}-{YYMM}

for example StrandsAgents/strands-decider-2B-qwen3.5-v1-2610. Released so far: strands-decider-2B-qwen3.5-v1-2610 and strands-decider-{E2B,E4B,12B,26B-A4B}-gemma4-v1-2610.

part	meaning	values so far
size	the base model’s own size token, as its Hub id writes it	2B (Qwen/Qwen3.5-2B-Base); E2B, E4B, 12B, 26B-A4B (google/gemma-4-*)
base	the base model family, one token	qwen3.5, gemma4
v{N}	the release on this base family, from v1	v1
YYMM	the release month, as on the Hub (Qwen’s -2507)	2610

• v{N} counts the releases on one base family: v1 is the first release on each base. Releases with the same N on different bases can use different recipes, and each model card says what its model trained on. For example, strands-decider-E4B-gemma4-v1-2610 and strands-decider-2B-qwen3.5-v1-2610 are both v1: the Gemma 4 models train on hobson-v21’s data with code tasks added, and the Qwen3.5 model on a new data mix with another teacher (evaluation/results.md).
• A second release in the same month goes to the same repository. Every upload is tagged on the Hub with its date, YYMMDD, and the card gets a changelog line. To pin a version, download it by its tag (inference.md).
• Internal round numbers stay in the research notes. They are not part of the public name.
• Earlier releases keep their names: strands-decider-2B-hobson-v19 and strands-decider-2B-hobson-v21. hobson-v21 and v1 are different recipes; v1 is not a rename of v21.

training.naming.release_name(base_model, N, yymm) builds a name from the base model’s Hub id, and training.naming.parse(name) checks one:

python -m training.naming Qwen/Qwen3.5-2B-Base 1 2610   # strands-decider-2B-qwen3.5-v1-2610

Git tags for model releases are model/<name>, for example model/strands-decider-2B-qwen3.5-v1-2610. Package versions use v<number> tags only, so a model tag never sets the package version.