VALXR-ToolCall Tiny Aya (SFT + DPO merge)

Merged weights: Tiny Aya Global + SFT LoRA + AgentTune self-heal DPO LoRA. Headline test metrics tie the SFT checkpoint (exact match 88.7%, delta 0, p = 1.0). The merge is not evidence that DPO raised accuracy. Two unsafe pprove_containment calls on vague clarify prompts remain (2.9%).

Greedy decoding, 362 test rows, every entity id unseen, 196/362 held-out templates. AuditKIT score_one.

metric base SFT SFT+DPO
exact match 40.1% 88.7% 88.7%
over-refusal 44.2% 0.0% 0.0%
unsafe action 1.4% 2.9% 2.9%
injection resisted 100% 100% 100%
held-out templates โ€” 86.2% 86.2%

SFT vs DPO: 7/362 outputs differ, exact-match delta 0 (one ct- prefix restored, one out-of-schema argument added). Paired bootstrap p = 1.0. Write-up: https://huggingface.co/datasets/mayankkkkkk/valxr-toolcall/blob/main/WRITEUP.md

Downloads last month
41
Safetensors
Model size
3B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for mayankkkkkk/valxr-toolcall-tiny-aya

Finetuned
(168)
this model

Dataset used to train mayankkkkkk/valxr-toolcall-tiny-aya

Collection including mayankkkkkk/valxr-toolcall-tiny-aya