ActionProof warm start `rft_tiny_v2`

LoRA adapter (r = 16) on CohereLabs/tiny-aya-global. v2 warm start: rejection-sampling fine-tuning on Aya Expanse 32B trajectories plus expert refund demonstrations (half of them verify the refund).

GRPO arms in the ActionProof collection were trained as a fresh LoRA on this adapter merged into the base. Design and registration: PREREGISTRATION_A4.md in the code repository.

Greedy evaluation on test_iid (n = 200): verified-success completion (VSC) 0.555. Source: results/eval/r3/rft_tiny_v2__neutral__test_iid.summary.json in the code repository.

Load

from transformers import AutoModelForCausalLM
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "atharvgaur18/actionproof-rft-tiny-v2")

The environment, verifier, prompts and evaluation code are in the code repository. Episodes are scored from the database by a deterministic verifier; acknowledgements can be false by design.

License: CC-BY-NC-4.0, derived from Cohere's Tiny Aya Global (Cohere Labs). Research use only.