Pinned dependency set cannot be installed as declared

## Summary

The dependency versions in `requirements.txt` cannot form a valid environment as declared. In particular, the file pins `numpy==1.11.1` while `transformers==5.9.0` declares `numpy>=1.17`.

This is still present on `main` at `a0b47382f08b108a17a39b343c4e58dbf2852db3`.

## Reproduction

On Windows 10 with Python 3.13.13 and pip 26.0.1:

```console
> python -m pip install --dry-run --ignore-installed --only-binary=:all: -r requirements.txt
Collecting lmdb==1.7.5
ERROR: Could not find a version that satisfies the requirement numpy==1.11.1
ERROR: No matching distribution found for numpy==1.11.1
```

PyPI metadata also shows that `transformers==5.9.0` requires `numpy>=1.17`, which directly conflicts with line 2 regardless of whether an old NumPy source build is available.

The other scientific-package pins (`pandas==0.14.1` and `scikit_learn==0.12.1`) likewise have no CPython 3.13 Windows wheels, but the NumPy/Transformers constraint conflict alone makes the full set unsatisfiable on every Python version.

## Expected behavior

The declared dependency set should resolve for a documented Python/CUDA platform so that the Quick Start can be reproduced.

## Impact

A fresh user cannot create the environment described by the repository, so training, evaluation, and regression validation cannot start from the published instructions.

## Suggested direction

- document the supported Python, CUDA, and platform versions;
- choose mutually compatible NumPy/Pandas/scikit-learn/Transformers versions;
- validate the resulting environment with a clean-install smoke check (and preferably a lock or constraints file).

## Why there is no PR in this report

Updating these pins requires a maintainer decision about the supported runtime and compatibility testing against the real training data/GPU stack. A speculative version bump would not be a complete or safely validated fix.

Start with requirements.txt and the repository's Quick Start instructions; reproduce the shown pip dry-run on a clean environment. Establish the supported Python, CUDA, and platform combination, then validate compatible dependency pins against the training stack. Done means the documented environment installs cleanly and the Quick Start can begin, ideally with a lock or constraints file.

Written by the indexing model from the issue text.