qwen3-asr-1.7b-ar-run6b-soup70-sip-v5-grpo2

sip-v5-grpo1 refined with a second round of reinforcement learning, MWER-style: for each training clip the beam search's 8 best transcripts are rewarded with 1 - WER against the label (Test Arabic ASR Benchmark normaliser), -1 for a repetition loop or text far longer than the label, and the model is pushed towards the better ones (advantages standardised within the clip's group), with a KL penalty to sip-v5-grpo1. LoRA r 16 on the language model's projections (merged; adapter in adapter/).

Prompts (~7,100 clips, 10 h): clips where sip-v5-grpo1 makes mistakes, plus 20% it gets right and 300 clips without speech, from in-domain Arabic call-centre recordings (private), EdAcc and Earnings-22 (accented English conversation, CC BY-SA 4.0).

Licence note: part of sip-v5's training text is CC BY-NC 4.0 and EdAcc / Earnings-22 are CC BY-SA 4.0; check that your use is compatible.

Evaluation (bf16, WER %; lower is better)

set	sip-v5	sip-v5-grpo1	sip-v5-grpo2
Audioluna call samples (20 recordings)	13.33	13.50	13.43
Test Arabic ASR Benchmark (505 segments)	9.99	9.59	10.39
Arabic segments	15.35	14.76	15.69
English segments	6.39	6.08	6.77
FLEURS ar_eg test (428)	5.25	5.06	5.61
In-domain call-centre test split	10.50	10.78	10.57
Clips without speech that get text (of 400)	107	104	101

Not recommended: this second (N-best) round is worse than sip-v5-grpo1 on the Test Arabic benchmark (+0.80) and FLEURS (+0.55) and no better on the call samples; use sip-v5-grpo1 instead.