OpenEuroLLM Function Calling

This repository is the public index for OpenEuroLLM function-calling data, post-training, evaluation, and research artifacts. Each software component remains in its own repository so it can preserve its upstream history and development workflow.

The controlled Qwen3.5 campaign and the TMAX reproduction are concluded. Their repositories are retained as historical source and infrastructure records, not as active training campaigns or validated capability results. Current work is tracked in Taskboard issue #220.

Available components

Component	Purpose	Release state
fcanalysis	Active toolkit for loading, normalizing, validating, and analyzing public function-calling datasets	Public active development
rollouts	Experimental code, attempt records, and selected trajectories for automatic task discovery and verification	Public experimental work
function-calling-post-training	Historical OLMo 3 and Qwen3.5 data preparation, training, and checkpoint-conversion source snapshots	Public historical record
Qwen3.5 controlled-campaign branch	Source snapshot for the concluded controlled campaign reported in #352	Concluded, no capability or mechanism result
OLMo 3 historical branch	Source snapshot for the historical single-training-seed OLMo 3 pilot	Historical, hypothesis-generating only
OLMo 3 result tag	Immutable source revision associated with the historical pilot	Tagged source, no checkpoint release
tmax-reproduction	Code-only record of the partial TMAX infrastructure reproduction reported in #351	Archived, no RL-performance result
bfcl	OpenEuroLLM snapshot of the Berkeley Function Calling Leaderboard evaluator	Public historical harness
tau2-bench	OpenEuroLLM snapshot of the tau2-bench evaluator	Public historical harness

Artifact availability

• No current OpenEuroLLM function-calling checkpoint release is available.
• The historical OLMo 3 and Qwen3.5 campaign checkpoints are not retained for publication.
• The historical processed five-source training dataset is not scheduled for release.
• fcanalysis is source-only. Its repository does not contain source datasets, generated training mixtures, model checkpoints, or benchmark outputs.

Machine-local evaluation harnesses, operational records, and private research material are outside the scope of this public index.

See components.yaml for a machine-readable component list.