# L1 embedding fine-tune — pinned deps.
#
# torch is intentionally NOT listed here: it ships in the CUDA base image's
# pip install step in the Dockerfile (pinned to a cu124 wheel) so we control the
# CUDA build explicitly rather than letting pip resolve a CPU/other-CUDA wheel.
# If you run this script outside the container, install torch yourself first.
#
# Version rationale:
# - sentence-transformers 5.x  : modern training stack — SentenceTransformerTrainer,
#                                SentenceTransformerTrainingArguments, MatryoshkaLoss,
#                                MultipleNegativesRankingLoss, DatasetDict round-robin
#                                multi-dataset training, batch_sampler="no_duplicates".
# - transformers 4.51.x        : minimum that includes the Qwen3 / Qwen3-Embedding
#                                architecture. Older transformers cannot load
#                                Qwen/Qwen3-Embedding-0.6B.
# - datasets 3.x               : Dataset / DatasetDict APIs used by the trainer.
# - accelerate 1.x             : required backend for the HF/ST Trainer loop.
# - huggingface_hub >=0.30,<1.0: push_to_hub(private=True) for the optional HF push.
#                                Lower bound is forced by transformers 4.51.3
#                                (requires hub>=0.30); upper bound <1.0 because the
#                                transformers 4.51 line caps below the 1.x hub. A
#                                hard ==0.27.1 pin is ResolutionImpossible here.
# - hf_transfer                : 3-5x faster shard downloads of the base model.
# - numpy <2 pin               : avoids ABI breaks against torch cu124 wheels built
#                                against numpy 1.x.

sentence-transformers==5.1.1
transformers==4.51.3
datasets==3.6.0
accelerate==1.6.0
huggingface_hub>=0.30.0,<1.0
hf_transfer==0.1.9
numpy==1.26.4
