data_factory_hard_v2_cnt_neonicon_supplement
收藏资源简介:
# tmp4_8 hard_v2 cnt / neonicon supplement `tmp4_8/` 是 `hard_v2` 的独立补充工作区,只负责本轮 `cnt + neonicon + neon_new` supplement,不改动 `tmp4_6/` 和 `tmp4_7/` 既有合同。 ## 目标 - 输出 31 个 derived supplement majors - 对外 `major/category` 使用 coarse 名字,不带 `_hard_` - 默认每个 derived major `50` 条 - 最终总量 `1550` 条 - 正式发布仓库:`Mrakas/data_factory_hard_v2_cnt_neonicon_supplement` - parent dataset repo:`Mrakas/data_factory_hard_v2` ## 主要 major - `magic_cot_cnt4_{L1,L1_5,L2,L3}` - `magic_cot_cnt5_{L1,L2}` - `cups_cot_cnt4_{L1,L1_25,L1_5,L2,L3}` - `cups_cot_cnt5_{L1}` - `magiccount_cot_cnt4_{L1,L1_5,L2,L3,L4}` - `magiccount_cot_cnt5_{L1,L2}` - `neonicon_cot_{L1,L1_5,L2}` - `neonicon_L1_cnt_{L1,L1_5,L2}` - `neon_new_{L1,L2,L3}` - `neon_new_cot_{L1,L2,L3}` ## 主要文件 - `hard_v2_cnt_neonicon_supplement_registry.py` - 锁定 31 个 supplement majors - 锁定 repo / manifest / source mapping / count override / word length ranges - `hard_v2_cnt_neonicon_supplement_common.py` - jobs / manifest / 相对路径 / dataset 公共逻辑 - `hard_v2_cnt_neonicon_supplement_rich_family.py` - 复用 round-rich cups / magic / magiccount - 实现带约束的 `neonicon` easing - 实现 `neon_new` / `neon_new_cot` - `build_data_factory_hard_v2_cnt_neonicon_supplement.py` - 单机 builder,支持 `--major/--majors` - `build_data_factory_hard_v2_cnt_neonicon_supplement_remote.py` - 服务器 bounded-queue orchestrator - `sanitize_data_factory_hard_v2_cnt_neonicon_supplement_questions.py` - 发布前清理 question 里的 `<think>` 包装提示 - `publish_data_factory_hard_v2_cnt_neonicon_supplement.py` - 上传到 ModelScope ## 产物 - `data_factory_hard_v2_cnt_neonicon_supplement_test.json` - `data_factory_hard_v2_cnt_neonicon_supplement_build_manifest.json` - `data_factory_hard_v2_cnt_neonicon_supplement_testset.zip` - `videos/<derived_major>/...` - `cases/<derived_major>/...` ## 远端命令 环境检查: ```bash /root/miniconda3/bin/python3 /data-tmp/ghx/tmp2/data_factory_v2/tools/check_generation_env.py --format text ``` 逻辑测试: ```bash /root/miniconda3/bin/python3 -m pytest -q \ /data-tmp/ghx/tmp2/data_factory_v2/tmp4_8/test_hard_v2_cnt_neonicon_supplement_registry.py \ /data-tmp/ghx/tmp2/data_factory_v2/tmp4_8/test_hard_v2_cnt_neonicon_supplement_scaling.py \ /data-tmp/ghx/tmp2/data_factory_v2/tmp4_8/test_hard_v2_cnt_neonicon_supplement_neonicon_constraints.py \ /data-tmp/ghx/tmp2/data_factory_v2/tmp4_8/test_hard_v2_cnt_neonicon_supplement_common.py \ /data-tmp/ghx/tmp2/data_factory_v2/tmp4_8/test_hard_v2_cnt_neonicon_supplement_path_helpers.py \ /data-tmp/ghx/tmp2/data_factory_v2/tmp4_8/test_hard_v2_cnt_neonicon_supplement_publish.py \ /data-tmp/ghx/tmp2/data_factory_v2/tmp4_8/test_hard_v2_cnt_neonicon_supplement_remote_plan.py ``` smoke build: ```bash /root/miniconda3/bin/python3 /data-tmp/ghx/tmp2/data_factory_v2/tmp4_8/build_data_factory_hard_v2_cnt_neonicon_supplement_remote.py --count-per-major 2 ``` 全量 build: ```bash /root/miniconda3/bin/python3 /data-tmp/ghx/tmp2/data_factory_v2/tmp4_8/build_data_factory_hard_v2_cnt_neonicon_supplement_remote.py --count-per-major 50 ``` 上传: ```bash MODELSCOPE_API_TOKEN=... /root/miniconda3/bin/python3 /data-tmp/ghx/tmp2/data_factory_v2/tmp4_8/publish_data_factory_hard_v2_cnt_neonicon_supplement.py \ --repo-id Mrakas/data_factory_hard_v2_cnt_neonicon_supplement \ --artifacts data_factory_hard_v2_cnt_neonicon_supplement_testset.zip \ --artifacts data_factory_hard_v2_cnt_neonicon_supplement_test.json \ --artifacts data_factory_hard_v2_cnt_neonicon_supplement_build_manifest.json \ --artifacts README.md ``` ## 路径合同 - 视频字段使用相对路径:`videos/<derived_major>/<file>.mp4` - `path == video_path` - 如有图片字段,继续使用相对路径:`images/<derived_major>/<file>.png` ## 发布约束 - question 文本里不要要求模型把推理写进 `<think>...</think>` - supplement 是完整独立子集,不原地修改 `Mrakas/data_factory_hard_v2` - 正式 build / test / sanitize / upload 只在服务器上执行



