Fix two issues with canonical Illumina evals

[ozan]

Hi there, I encountered a couple of small issues with the ground truth for the Illumina canonical evals.

Firstly for the batch key choice, the correct answer I believe should be A (sample) since the sample key does in fact exist in the data. To repro:

latch cp latch://159007346.node /tmp/illumina_drg_main.h5ad
python3 -c "import anndata; print('sample' in anndata.read_h5ad('/tmp/illumina_drg_main.h5ad').obs.columns)"

... which shows as True. I confirmed that the data in sample is meaningful (3 samples, realistic distribution) so it seems like just a typo in the ground truth.

Secondly there's a minor (within tolerance) issue with the QC eval. The prompt indicates a threshold of n_genes > 1000 but the ground truth values are calculated as if the prompt is >= 1000 (greater than or equal). I've updated the prompt rather than the GT values but you could equally update these from 19,270 / 11,521 / 7,749 -> 19,259 / 11,518 / 7,741.

To repro:

"""
Data file downloaded via `latch cp latch://158807664.node /tmp/illumina_drg_qc.h5ad`
"""
import anndata, numpy as np, scipy.sparse as sp

X = anndata.read_h5ad("/tmp/illumina_drg_qc.h5ad").X
n = np.array((X > 0).sum(axis=1)).flatten() if sp.issparse(X) else (X > 0).sum(axis=1)
print(f"n_genes >  1000: {(n > 1000).sum()}  (eval instructions)")
print(f"n_genes >= 1000: {(n >= 1000).sum()}  (matches GT of 19270)")

Out:

n_genes >  1000: 19259  (eval instructions)                                                                                         
n_genes >= 1000: 19270  (matches GT of 19270)