围绕 GPT-6 Astra 搭微调数据闭环的六个仓库

Promix @Promixyz原推发布于 2026.10.17

在 X 上打开原推

better models will not fix a bad dataset 6 GitHub repos build a finetune loop around GPT 6 Astra 01 Distilabel ▸ https://github.com/argilla-io/distilabel → generates synthetic instruction and preference data 02 Argilla ▸ https://github.com/argilla-io/argilla → curates human feedback on the training set 03 cleanlab ▸ https://github.com/cleanlab/cleanlab → finds label errors before you waste a training run 04 Unsloth ▸ https://github.com/unslothai/unsloth → speeds up the finetune kernels on the cleaned set 05 TRL ▸ https://github.com/huggingface/trl → runs SFT and DPO once the data is ready 06 Axolotl ▸ https://github.com/axolotl-ai-cloud/axolotl → keeps the full train recipe in one YAML file how to replicate 1 generate candidates in Distilabel 2 review and label in Argilla 3 drop bad rows with cleanlab 4 train with Unsloth plus TRL 5 lock the recipe in Axolotl so the run is repeatable Most teams buy a newer model when the dataset is the bottleneck Fix the data loop once and every model you plug in gets cheaper wins

这条是引用着另一条写的：x.com/lummox_eth/status/2111100800625787154