You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These files contain the full text of the Decision Index frozen suite, including GPQA and HLE questions whose authors ask that they not be posted in plain text. Use them to verify the scores; do not train on them.
Log in or Sign Up to review the conditions and access this dataset content.
Decision Index results β Surogate Rune 26B-A4B
The complete Decision Index run of
Surogate Rune 26B-A4B (rune-26b-a4b-JD-Q6_K.gguf),
served by surogate through its decisions endpoint.
| Decision Index | balanced skill | breadth skill | requests |
|---|---|---|---|
| 57.24 | 43.24 | 41.91 | 132,422 of 132,422 completed (complete: true) |
Files
runs/rune-26b-a4b-jd-q6k/
results.jsonl.gz: every row as written by the kit's runner.scores.json,index.json,benchmark-summary.json: the kit'sscoreoutput for that file.shards/shard-*/environment.json,status.json: the runner's records for each of the 32 shards.harness/: the engine adapter (decision_index_surogate_engine_v1.py) and the launcher that ran the shards.
How it was run
- Kit: commit
52a6989, frozen suite verified by hash. - Engine adapter:
decision_index_surogate_engine_v1:SurogateEngine. It forwards each row unmodified as onePOST /api/alpha/decisionsrequest: one shared prefill of the state, then one suffix per question. It truncates nothing, drops no option and adapts no prompt per benchmark. - Server: surogate commit
19ff7d38(release v1.5.2), one server per GPU:surogate serve rune-26b-a4b-JD-Q6_K.gguf --max-model-len 32768 --max-num-seqs 8 --kv-capacity 32768. Thinking is off; the readout is the option-letter distribution at the first generated position. - Hardware: 8 Γ NVIDIA RTX 5090 with power capped at 400 W, 8 servers, 32 runner shards (4 per server).
The shard results were concatenated and scored once with the kit's
score. - Outcome: 132,421 rows
okand 1error. The error is an HLE request refused by the endpoint's template-boundary guard, and it scores as wrong. There are no unsupported rows and nothing was truncated. Choice questions up to 255 options are supported. - Latency: median 492.8 ms and p95 7,137 ms. These are HTTP wall times per row under full 32-worker load.
- Run history: the first attempt was stopped at about 2,300 rows by an engine admission bug, which is fixed in
19ff7d38. The run was resumed with the kit's own checkpoint/resume. A final resume pass re-ran 22 rows that had exhausted retries on queue-full (429) responses under load. The fix concerns request admission only; it does not change any output.
Masked local paths
To remove personal data, local filesystem paths in our own fields were rewritten before upload, and nothing else was changed:
- the served GGUF's local path became
/tmp/rune-26b-a4b-JD-Q6_K.gguf; - home-directory paths in tracebacks and records became
/workspace/....
The suite text and all scored content are byte-identical. Re-scoring the published results.jsonl.gz with the kit
reproduces the published scores.json, index.json and benchmark-summary.json exactly, apart from the generation
timestamp.
The served GGUF has the same tensors as the published rune-26b-a4b-JD-Q6_K.gguf. The published file's chat template
was later changed only so that thinking is off by default, which the decisions endpoint already enforces. In
harness/launch.sh, one variable name and the comments about GPUs shared with another workload on our machine were
generalised.
- Downloads last month
- 48