Trained Verifier Models

updated Aug 12

Surrogate code verifiers across three model sizes trained using multiple different algorithms as described in the Aletheia paper

Upvote

Aletheia-Bench/GRPO-Think-1.5B-16k

Text Generation • 2B • Updated Aug 25 • 27

Note Our flagship verifiers, trained using a GRPO-style algorithm with a 16k generation limit
Aletheia-Bench/GRPO-Think-7B-16k

Text Generation • 8B • Updated Aug 25 • 26

Note Our flagship verifiers, trained using a GRPO-style algorithm with a 16k generation limit
Aletheia-Bench/GRPO-Think-14B-16k

Text Generation • 15B • Updated Aug 25 • 19

Note Our flagship verifiers, trained using a GRPO-style algorithm with a 16k generation limit
Aletheia-Bench/GRPO-Think-1.5B-4k

Text Generation • 2B • Updated Aug 25 • 18

Note A variant of the flagship verifiers trained with a 4k generation limit
Aletheia-Bench/GRPO-Think-7B-4k

Text Generation • 8B • Updated Aug 25 • 27

Note A variant of the flagship verifiers trained with a 4k generation limit
Aletheia-Bench/GRPO-Think-14B-4k

Text Generation • 15B • Updated Aug 25 • 24

Note A variant of the flagship verifiers trained with a 4k generation limit
Aletheia-Bench/GRPO-Think-1.5B-8k

Text Generation • 2B • Updated Aug 25 • 44

Note A variant of our flagship verifiers, trained with a 8k generation limit
Aletheia-Bench/GRPO-Think-7B-8k

Text Generation • 8B • Updated Aug 25 • 30

Note A variant of our flagship verifiers, trained with a 8k generation limit
Aletheia-Bench/GRPO-Think-14B-8k

Text Generation • 15B • Updated Aug 25 • 34 • 1

Note A variant of our flagship verifiers, trained with a 8k generation limit
Aletheia-Bench/GRPO-Instruct-1.5B

Text Generation • 2B • Updated Aug 25 • 26

Note A GRPO training ablation that does not generate thinking traces
Aletheia-Bench/GRPO-Instruct-7B

Text Generation • 8B • Updated Aug 25 • 27

Note A GRPO training ablation that does not generate thinking traces
Aletheia-Bench/GRPO-Instruct-14B

Text Generation • 15B • Updated Aug 25 • 23

Note A GRPO training ablation that does not generate thinking traces
Aletheia-Bench/DPO-Think-1.5B

Text Generation • 2B • Updated Aug 25 • 25

Note A verifier trained completely offline using DPO
Aletheia-Bench/DPO-Think-7B

Text Generation • 8B • Updated Aug 25 • 22

Note A verifier trained completely offline using DPO
Aletheia-Bench/DPO-Think-14B

Text Generation • 15B • Updated Aug 25 • 45 • 2

Note A verifier trained completely offline using DPO
Aletheia-Bench/BatchOnline-GRPO-1.5B

Text Generation • 2B • Updated Aug 25 • 36

Note A variant of our flagship verifiers, where the generation policy is synced every 4 gradient updates
Aletheia-Bench/BatchOnline-GRPO-7B

Text Generation • 8B • Updated Aug 25 • 29 • 1

Note A variant of our flagship verifiers, where the generation policy is synced every 4 gradient updates
Aletheia-Bench/BatchOnline-GRPO-14B

Text Generation • 15B • Updated Aug 25 • 26 • 1

Note A variant of our flagship verifiers, where the generation policy is synced every 4 gradient updates
Aletheia-Bench/RAFT-1.5B

Text Generation • 2B • Updated Aug 25 • 21

Note A verifier trained using on-policy rejection sampling on only positive samples
Aletheia-Bench/RAFT-7B

Text Generation • 8B • Updated Aug 25 • 28

Note A verifier trained using on-policy rejection sampling on only positive samples
Aletheia-Bench/RAFT-14B

Text Generation • 15B • Updated Aug 25 • 23

Note A verifier trained using on-policy rejection sampling on only positive samples
Aletheia: What Makes RLVR For Code Verifiers Tick?

Paper • 2601.12186 • Published Jun 1 • 1

Upvote