tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

honor skip prepare for rl

Browse Source
This commit is contained in:
Wing Lian
2025-02-02 23:39:10 -05:00
parent 54b0d3d0e8
commit 704ddd6ff1
1 changed files with 19 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
src/axolotl/utils/data/rl.py
View File

@@ -163,23 +163,26 @@ def load_prepare_preference_datasets(cfg):
# "prompt", "chosen" and "rejected" already preprocessed
split_datasets[i] = data_set
drop_long = partial(
drop_long_rl_seq,
rl=_cfg.rl,
tokenizer=tokenizer,
sequence_len=cfg.sequence_len,
)
if not cfg.skip_prepare_dataset:
drop_long = partial(
drop_long_rl_seq,
rl=_cfg.rl,
tokenizer=tokenizer,
sequence_len=cfg.sequence_len,
)
prior_len = len(split_datasets[i])
split_datasets[i] = split_datasets[i].filter(
drop_long,
num_proc=cfg.dataset_processes,
load_from_cache_file=not cfg.is_preprocess,
desc="Dropping Long Sequences",
)
dropped = prior_len - len(split_datasets[i])
if dropped:
LOG.warning(f"Dropped {dropped} long samples from dataset index {i}")
prior_len = len(split_datasets[i])
split_datasets[i] = split_datasets[i].filter(
drop_long,
num_proc=cfg.dataset_processes,
load_from_cache_file=not cfg.is_preprocess,
desc="Dropping Long Sequences",
)
dropped = prior_len - len(split_datasets[i])
if dropped:
LOG.warning(
f"Dropped {dropped} long samples from dataset index {i}"
)
combined_datasets = concatenate_datasets(split_datasets)
combined_datasets = combined_datasets.shuffle(seed=cfg.seed)