CASE EVIDENCE / A RECORD
AIR-hl (research group) 公开的多模态生成与理解案例，来源为 公开代码库，复核于 2026-06-27T01:40:00Z。
原始记录：AIR-hl RL training with DAPO on DeepSeek-V3.1-Terminus using FP8 quantization
Chinese Brief
AIR-hl (research group) 公开的多模态生成与理解案例，来源为 公开代码库，复核于 2026-06-27T01:40:00Z。 Model Atlas 将它标记为 A 类证据，因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。 Model Atlas 不把 benchmark、教程、发布说明或集合页包装成真实案例。
任务
这是一个围绕多模态内容处理的真实任务，公开材料可以回溯到具体使用者和具体产物。 原始资料写作：Training DAPO (Direct Alignment from Preference Optimization) on DeepSeek-V3.1-Terminus with FP8 format for reinforcement learning to improve reasoning capabilities. The training runs on a 24-node cluster with 8 NVIDIA …
公开材料提供公开代码、README 或项目配置，可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作：Encountered a dimension mismatch error (576 must be a multiple of block_size0: 128) during FP8 weight quantization in the verl training pipeline. The issue was filed on 2026-01-07 and received 14 comments, indicating ac…
DeepSeek V3.1 Terminus 在该案例中承担多模态内容处理相关的生成、分析、编排或实现角色。 原始资料写作：DeepSeek-V3.1-Terminus serves as the base model for RL training. Its 671B-parameter MoE architecture with Mixture-of-Experts design enables efficient training at scale. The model's improved language consistency and agen…
当前判断基于公开材料；若产物下线、仓库变更或模型参与比例仅来自作者自述，需要在引用前重新复核。 原始资料写作：The user encountered a technical error during FP8 training, indicating potential compatibility issues with certain quantization formats. The issue is actively being debugged. The model's large size (671B params) require…