You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access to these generations is reviewed manually. Please state your name, affiliation, and how you intend to use them.

Log in or Sign Up to review the conditions and access this dataset content.

sft_gene_pathway_v7_6_100k_relabel_9b — generations

Every raw generation behind the reported scores for wanglab/sft_gene_pathway_v7_6_100k_relabel_9b, at both decodings, released so the numbers can be recomputed rather than taken on trust.

The repo name is the run name, so it maps 1:1 onto the source tree:

experiment leaf experiments/sft_gene_pathway_v7_6_100k/9b_full_relabel/
training run sft_gene_pathway_v7_6_100k_relabel_9b_stage2
checkpoint scored sft_gene_pathway_v7_6_100k_relabel_9b_converted, stage 2 last.ckpt (step 14,480)
source dirs evals_bench/greedy/, evals_bench_sc/maj10/
scoring set benchmark_genetic_v1

Contents

Scored set is benchmark_genetic_v127,067 rows / 1,319 perturbations.

split decoding rows files
greedy/ greedy, 1 sample per row 27,067 5
maj10/ temperature 1, 10 samples per row 270,670 5

One file per source dataset in each:

dataset rows perturbations
xaira_orion 9,929 598
nadig 8,708 292
marson 4,579 240
xaira_pisces 2,186 110
kaufman 1,665 77

Schema

JSONL, one record per generation:

field type meaning
idx int row index into the benchmark; join on this, never on position
prompt str the rendered prompt as fed to the model
generation str the model's reasoning trace and boxed answer
ground_truth str the reference completion, carrying the true label in \boxed{}
timestamp str ISO-8601 generation time

Two things to know before parsing:

Join on idx, not on line order. Generation is emitted in dataset order, so a positional join is correct for a file written in one pass — but a run requeued after an interruption replays a block, and the file then carries duplicate idx out of order. The last write of an idx wins. This is not hypothetical; it happened during this project and a positional join would have silently scored every row after the tear against the wrong benchmark row.

prompt is the embedded-prompt form. Cell state enters as a projected embedding rather than text, so the prompt field carries padding tokens where the cell block sits. It is the literal model input, not a human-readable rendering.

Scoring

The label is the direction inside \boxed{}. The metric of record is committed accuracy — accuracy over rows where the model commits to upregulated or downregulated — reported as the per-perturbation macro-average, with the commit rate beside it. Models abstain at different rates and that is a property worth reporting, not a nuisance to average away.

For maj10/, take the modal parse over the ten samples per idx.

Always quote the constant-direction control alongside any number: always answering that dataset's majority direction scores 58.6 pooled, and per-dataset from 51.3 to 65.5. It is the bar to clear.

Provenance

Generated from stage 2's last.ckpt (step 14,480) of the 9b_full_relabel run, at 100% row coverage on all ten dataset × decoding cells as of 2026-08-31.

Any maj@10 figure for this model quoted before that date is superseded: an earlier table was computed while nadig and xaira_orion were still filling at 89–91%, and the missing rows were not neutral — completing them moved xaira_orion 76.4 → 74.0 and nadig 61.1 → 59.7, i.e. the partial figures ran ~1.5–2.5 points optimistic on exactly the two incomplete datasets.

License

Generations from a model derived from Qwen3.5-9B, itself trained on Claude Opus 5 outputs. Use is subject to the terms governing both.

Downloads last month
22