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

Sieve-2B: Decision Index 0.2.1 results

The complete run of Sieve-2B on the frozen suite of the Decision Index. The run used the official kit apolinario/decision-index at 87d4650 (Version 0.2.1), and the kit's own scorer produced the scores.

Decision Index 0.2.1: 21.74 (raw index 39.91, breadth 20.55). All 150,317 scoreable requests have results: 150,315 answered, 2 unsupported (options with identical text) and 0 errors. The runner file also contains the 442 official scoring exclusions, so it has 150,759 rows in total.

Not on the leaderboard yet. This is a self-reported run. 21.74 is the score we expect, not a confirmed leaderboard result, until the maintainers review the submission and add it.

area	skill (×100)
Knowledge & Reasoning	10.6
Language Understanding	19.7
Retrieval & Classification	28.5
Tools & Automation	35.3
Arts & Human Taste	17.8

Files (runs/Sieve-2B/)

file	contents
scores.json	the kit's scores: "complete": true, edition 0.2.1, index, areas, per-benchmark values
index.json, benchmark-summary.json	per-benchmark raw, skill and coverage; native metrics and counts
environment.json	the runner's environment, plus an execution block describing the sharded run
status.json	final status of the merged run (complete, 150,759 rows: 150,734 ok, 25 unsupported)
results.jsonl.gz	one result per request: run id, status, the model's answers with a probability for every option, and timings
overlap_run_ids.json	every suite run_id that shares an input with the training data or the temperature pool

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; every row keeps its payload_sha256. Scoring this file with python -m decision_index score --results runs/Sieve-2B/results.jsonl.gz reproduces scores.json exactly, apart from its generation time. The run was recorded in a folder named sieve-2b-full and was re-scored after the folder was renamed Sieve-2B, so scores.json names the run Sieve-2B. Every score is unchanged.

How the run was made

• Suite: rebuilt with suite rebuild and staged with suite import. The rows are byte-identical to the lab's (selected-rows uncompressed sha256 b2b56d6f…, added-rows 7429f3c9…). HLE was fetched with an account that accepted its terms. Raw suite inputs are not redistributed.
• Model: sthanika-ai/sieve-2b at revision 5aa4d83a4096bcb6d0e0ae75738a225a8fad4b1e. That is a LoRA adapter plus a scalar scorer head on Qwen/Qwen3.5-2B-Base, truncated to its first 16 of 24 blocks: 1.43B parameters run, 4.69M trained. The calibrated temperature (1.7144) is stored in config.json. Published files: adapter sha256 76e7c562…, head sha256 1805086a….
• Engine: sieve_engine:SieveEngine, published in the model repo at the same revision (sha256 46fb3287…, also recorded in environment.json).
  • The run itself used an earlier copy of the same engine that differed only in how it locates the model files (it now also loads by repo id). Loaded by repo id with the run's options, the published repo reproduces the recorded probabilities bit-for-bit (100 of 100 re-run requests, A100).
  • Options: max_batch_tokens=49152, mem_fraction=0.16. With default options the answers are identical and probabilities differ by up to 3e-2 (bf16 batching).
  • The state and the question (with every option listed, in an order fixed by a hash of each option's own text) are encoded once; the cache is copied once per option and each option continues from its own copy, so no option sees another and reordering the options changes nothing. A request's questions are scored together, grouped by length.
  • noul is answered as a 2-option choice over its own criteria descriptions, or plain "No." / "Yes." when none are given.
  • Nothing is truncated, no option is dropped, and no prompt is changed per benchmark.
  • Declared limits: state + question + the longest option within 32,768 tokens; no option-count limit. No request reached the limit. Questions whose options have identical text are refused as Unsupported.
• Hardware: 2× NVIDIA A100 80GB PCIe, shared with other users' jobs; bf16. torch 2.8.0+cu128, transformers 5.17.0, peft 0.21.0, flash-linear-attention 0.5.2, causal-conv1d 1.7.0.
• Execution: the suite was split round-robin into 24 shards of exactly the rows pipeline runs and scored by 7 concurrent processes (3 on one GPU, 4 on the other). The 24 results.jsonl files were concatenated with one record per run_id (no duplicates, no retries needed). Every shard ran the same engine, options, model and runner.
• Local latency: single process, one request at a time, bf16, idle A100: median 85 ms, p95 223 ms (400 uniformly sampled suite requests). The latency_ms in scores.json (median 143 ms) was recorded during the shared 7-process run. Neither is the maintainers' RTX PRO 6000 measurement.

Overlap with the training data

Every training row and the temperature-fitting pool were checked against every suite row (inputs only: a shared 13-word run, or a whole 5-12-word field contained in the other side). 33 of the 150,317 scored rows (0.02%) share an input with the training data: BANKING77 10 (short customer queries that occur word-for-word in BANKING77's own train and test splits), HoVer 9 and RAGTruth 6 (shared source sentences with unrelated labels), ANLI 2, SATA-Bench 2, iSarcasmEval 2, BRIGHT 1, MMLU 1 (not in the index). No benchmark's test set was used for training. The temperature pool (used only to fit the single temperature, which never changes a decision) shares inputs with 85 suite rows, mostly ContractNLI clauses, CLINC150 utterances and MuSR. Every run_id is listed in runs/Sieve-2B/overlap_run_ids.json.