Qwen3.5-9B Tajik/EN Tool-Calling (SFT + GRPO)
How to get access: please fill in the short form at saidzoda.com/access. We approve requests after we receive the form.
GRPO-refined tool-caller (reward = XML tool-call scorer) on top of the SFT model. Emits Qwen XML**
tool calls (<tool_call><function=name><parameter=k>v</parameter></function></tool_call>).
- Data: ~132k tool-call convs (tg+en, BFCL-derived) + ~44k Tajik instruct replay (preserves reasoning). 1 epoch, LoRA r32.
- Tool-calling accuracy (n=100/lang): TG mean-reward 0.631 (fn-name 0.71, full-call 0.43) · EN 0.819 (fn-name 0.90, full-call 0.60). Overall 0.725 (vs SFT 0.652; GRPO reward 0.43->0.69).
Gated (manual). Saidzoda Lab.