feat(megatron-bridge): SFT-masked data support in distillation - #2113

Merged

yueshen2016 merged 11 commits into

mainfrom

feat/mbridge-sft-distillation

Aug 13, 2026

Commits on Aug 13, 2026

feat(megatron-bridge): SFT-masked data support in distillation
yueshen2016
committed
14ccdf0
fix(megatron-bridge): let --sft run without pretraining --data_paths; document verbatim SFT format
yueshen2016
committed
fb1c84d
fix(megatron-bridge): address review on the SFT distillation path
yueshen2016
committed
cf733d5
fix(megatron-bridge): compare tokenizer mappings, not vocab sizes, and note the appended EOS
yueshen2016
committed
f015f01
feat(megatron-bridge): warn when --sft data lacks a BOS the tokenizer adds at inference
yueshen2016
committed
8057060
fix(megatron-bridge): enforce the verbatim SFT contract in the tokenizer, not just the dataset
yueshen2016
committed
44a2f42
fix(megatron-bridge): truncate SFT prompts from the left, and relax the tokenizer check
yueshen2016
committed
7baef79
fix(megatron-bridge): check logits width, not tokenizer size, and reject a stray --sft_dataset_root
yueshen2016
committed
34bf2fe
fix(megatron-bridge): address review on distillation vocabulary and SFT input
yueshen2016
committed
9467589
fix(megatron-bridge): simplify the tokenizer check, trim comments, add changelog
yueshen2016
committed
9fd399a
fix(megatron-bridge): raise on tokenizer failures and add BOS via the framework
yueshen2016
committed
8fe9543