sthanika-ai/Sieve-2B-v3-decision-index-results

Sieve-2B-v3: Decision Index results

The complete run of Sieve-2B-v3 on the public suite of the Decision Index 0.3, made with the official kit apolinario/decision-index at f5a6e4f (0.3); the kit's own scorer produced the scores.

Leaderboard	not on the leaderboard yet
This run	Decision Index 0.3: public index 53.51 (raw index 65.39, breadth 51.09)
Requests	all 140,178 scoreable requests answered: 0 unsupported, 0 errors

The Full score ranks the board: the public benchmarks count 20%, the maintainers' private tests of the same skills 50% and their private tasks from new domains 30%.

area	skill (×100)
Knowledge & Reasoning	30.9
Language Understanding	49.3
Retrieval & Classification	66.5
Tools & Automation	80.5
Arts & Human Taste	47.6

Files (runs/Sieve-2B-v3/)

file	contents
scores.json	the kit's scores: "complete": true, edition 0.3, index, areas, per-benchmark values
index.json, benchmark-summary.json	per-benchmark raw, skill and coverage; native metrics and counts
shard-0/ … shard-5/	the runner's environment and final status for each sixth of the run (23,436 or 23,437 requests each, all ok)
results.jsonl.gz	one result per request: run id, status, the model's answers with a probability for every option, and timings
RUN_NOTES.md	how the run was made, step by step, with checksums

results.jsonl.gz is in the runner's --compact form. The payload field (benchmark text) and the raw_output field are left out, because several suite sources do not allow their text to be republished.

Scoring this file with python -m decision_index score --edition 0.3 --engine sieve.decision_index_engine:SieveEngine --results runs/Sieve-2B-v3/results.jsonl.gz reproduces scores.json, index.json and benchmark-summary.json exactly, apart from their generation time.

How the run was made

• One complete 0.3 run (2026-10-15), on two NVIDIA A100 80GB PCIe GPUs. The 140,620 rows of the 0.3 suite were split into six shards of 23,436 or 23,437 rows, balanced by the suite's proxy token counts. Three runners shared each GPU, and the six results.jsonl files were concatenated. The file has 140,620 rows; the 442 official exclusions leave 140,178 scored.
• Suite. suite-0.3 from the kit (rows sha256 b2b56d6f…, added rows 7429f3c9…, rebuilt GSM8K 75933489…, all in scores.json). Raw suite inputs are not redistributed.
• Model. sthanika-ai/Sieve-2B-v3, a full fine-tune of Qwen/Qwen3.5-2B at 15852e8c (its 24 decoder layers, stored as a PEFT adapter) plus a pointer head. Its head.pt stores a global temperature (0.968) and one temperature per question category, 25 in all. The run loaded adapter sha256 6a33f22c… and head sha256 7ef0f311…; both are recorded in every environment.json and match the published files.
• Engine. sieve.decision_index_engine:SieveEngine from github.com/sthanika-ai/Sieve (0.3.0, which applies the per-category temperatures), with --option model=<local copy of Sieve-2B-v3>, --rows <shard> and --compact.
  • The adapter is merged into the bf16 backbone, the pointer head runs in fp32, and the path is eager.
  • Nothing is truncated, no option is dropped, and no prompt is changed per benchmark.
  • Declared limits: a 65,536-token state, a 32,768-token question and up to 255 choice options. Every request fit within them. The model was trained on states and questions of up to 16,384 tokens each.
• Software. Torch 2.8.0+cu128, transformers 5.17.0, peft 0.21.0, flash-linear-attention 0.5.2, triton 3.4.0 and causal-conv1d 1.7.0.
• Local latency (not the maintainers' measurement). In-process wall time, eager, over all 140,620 requests: the median is 59 ms, the mean 143 ms, the 80th percentile 115 ms and the 95th percentile 329 ms. Six engines (three per GPU) ran at once, next to other users' GPU jobs, so these are upper bounds.