Laya Needle V10: multilingual-full

A fine-tune of convaiinnovations/laya (revision 7b928d8), Multilingual (mmBERT-base encoder) checkpoint, as a tool-call decision model for the Needle V10 task. It was trained on candidate set full (all 36 V10 decision tools).

Training data: the Needle V10 tool-call corpus (1,767 JEV-assisted/hard-weighted training rows, 556-row family-disjoint validation split). The corpus is private and not published.
Recipe: upstream RLCD plus one-hot cross entropy, AdamW (encoder LR 2.5e-5, head LR 1e-4), BF16. Checkpoint chosen on the validation split only.
Result: negative. Scored 352/360 curated, 47/100 adversarial on its own candidate arm, and noticeably worse across arms. It did not beat the repaired Needle V10 (57/100 adversarial) and was not deployed. Published as a research artifact only.

Files are the best/ checkpoint as written by the Laya trainer (model.safetensors, encoder/, tokenizer/, rl_agent_config.json).

Downloads last month: 10

Safetensors

Model size

0.3B params

Tensor type

F16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for turnercore/laya-needle-v10-multilingual-full

Base model

convaiinnovations/laya

Finetuned

(191)

this model