# Training-only dependencies for CounterFeint Investigator GRPO fine-tuning.
#
# Kept in a SEPARATE file from requirements-dev.txt so:
#   * `pip install -e .` (runtime) stays slim — no torch/transformers in
#     the FraudArena Docker image (~5 GB saved).
#   * `pip install -r requirements-dev.txt` (CI tests + matplotlib) does
#     not pull a full GPU stack onto every laptop.
#   * `pip install -r requirements-train.txt` is the single command you
#     run on the Colab T4 / A100 node before starting GRPO training.
#
# Pinned to versions known-good with TRL's GRPOTrainer at the time of
# the hackathon. If TRL has shipped a newer release onsite, prefer that
# and update the version pins here in the same commit.

# Core PyTorch + HuggingFace stack
torch>=2.3.0
transformers>=4.46.0
accelerate>=1.1.0
datasets>=3.0.0

# PEFT (LoRA / QLoRA adapters) + 4-bit quantisation
peft>=0.13.0
bitsandbytes>=0.44.0

# TRL ships GRPOTrainer; >=0.12 is when it stabilised.
trl>=0.12.0

# safetensors is implicit but pin it to avoid torch-side warnings
safetensors>=0.4.5

# CLI helpers used by the notebooks
huggingface_hub>=0.26.0

# Reuse the same eval-time deps so the final cell (`run_before_after`)
# can render eval_plot.png from the training notebook.
matplotlib>=3.8.0

# Allow `asyncio.run()` inside Jupyter / Colab cells. CounterFeint's
# inference.run_three_agent_episode is a sync wrapper around an async
# core (arun_three_agent_episode); without nest_asyncio it raises
# "asyncio.run() cannot be called from a running event loop" the moment
# we collect rollouts from a notebook.
nest_asyncio>=1.6.0