From ab818c8e2f8a72ab541ba385c6bb2d4445adcdfa Mon Sep 17 00:00:00 2001 From: ruit Date: Thu, 29 Oct 2026 20:16:52 -0700 Subject: [PATCH] docs: point the AVQA reference at the dataset the recipe actually loads https://mn.cs.tsinghua.edu.cn/avqa now answers 403 and redirects to a webvpn gateway, so the docs linkcheck fails on main: { "filename": "guides/grpo-audio.md", "lineno": 18, "status": "broken", "uri": "https://mn.cs.tsinghua.edu.cn/avqa", "info": "403 Client Error: Forbidden for url: https://webvpn.tsinghua.edu.cn/..." } The 3B recipe does not load anything from that host. It pulls the pre-processed split from gijs/avqa-processed on the Hub, which the same section already names two sentences later. Point the recipe table there, matching how the 7B and 30B rows link their datasets. The prose mention keeps the dataset's full name but drops its link, since the sentence it sits in already links the processed dataset and the R1-AQA paper; a third link to the same place would add nothing. Co-Authored-By: Claude Opus 5 (1M context) Signed-off-by: ruit --- docs/guides/grpo-audio.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/docs/guides/grpo-audio.md b/docs/guides/grpo-audio.md index f5702712e8d..b76048cdd5a 100644 --- a/docs/guides/grpo-audio.md +++ b/docs/guides/grpo-audio.md @@ -15,7 +15,7 @@ NeMo RL ships three recipes out of the box, but the pieces are independent and c | Recipe | Model | Default dataset | Config | | --- | --- | --- | --- | -| 3B (R1-AQA reproduction) | Qwen2.5-Omni-3B | [AVQA](https://mn.cs.tsinghua.edu.cn/avqa) | `examples/configs/audio_grpo_3B_megatron.yaml` | +| 3B (R1-AQA reproduction) | Qwen2.5-Omni-3B | [AVQA](https://huggingface.co/datasets/gijs/avqa-processed) | `examples/configs/audio_grpo_3B_megatron.yaml` | | 7B (StrongAC + Gemini CoT) | Qwen2.5-Omni-7B | [Harland/AudioMCQ-StrongAC-GeminiCoT](https://huggingface.co/datasets/Harland/AudioMCQ-StrongAC-GeminiCoT) | `examples/configs/recipes/vlm/vlm_grpo-qwen2.5-omni-7b-audiomcq-1n8g-megatron.v1.yaml` | | 30B MoE (Qwen3-Omni) | Qwen3-Omni-30B-A3B-Instruct | Harland/AudioMCQ-StrongAC-GeminiCoT | `examples/configs/recipes/vlm/vlm_grpo-qwen3-omni-30ba3b-audiomcq-4n8g-megatron.v1.yaml` | @@ -25,7 +25,7 @@ The 3B recipe accepts the AudioMCQ dataset via a CLI override (no new YAML neede ### AVQA -The [Audio-Visual Question Answering (AVQA)](https://mn.cs.tsinghua.edu.cn/avqa) dataset is the original training corpus from the [R1-AQA paper](https://arxiv.org/abs/2503.11197). NeMo RL exposes it under the registry key `audiomcq`'s sibling, `dataset_name: avqa`, and pulls the pre-processed split from [`gijs/avqa-processed`](https://huggingface.co/datasets/gijs/avqa-processed) on the Hub. AVQA has native `train` and `validation` splits, so the 3B recipe references both directly. +The Audio-Visual Question Answering (AVQA) dataset is the original training corpus from the [R1-AQA paper](https://arxiv.org/abs/2503.11197). NeMo RL exposes it under the registry key `audiomcq`'s sibling, `dataset_name: avqa`, and pulls the pre-processed split from [`gijs/avqa-processed`](https://huggingface.co/datasets/gijs/avqa-processed) on the Hub. AVQA has native `train` and `validation` splits, so the 3B recipe references both directly. ### AudioMCQ-StrongAC-GeminiCoT