Datasets:

horus14-wq
/

frontiercs2-t3

question stringclasses 48 values	text stringlengths 300 1.36M	title stringclasses 612 values	source stringlengths 52 96
# Approximate-MIPS Query Stopping `frontier-submit train` returns immediately after acceptance with a verification ID and the authoritative count of remaining train calls. It does not wait for verification. Use `frontier-submit status <verification_id>` for one status read; once verification completes, it includes you...	Now I understand the setup. `should_stop` is called inside the while loop in knnF. The loop continues while `cnt < beta*MaxNum` where `beta=1.0f` and `MaxNum = hash.parti.nums[chunks]` — the number of vectors in the partition. So the fixed-candidate-budget starter policy is basically "stop once all candidates in the pa...	approximate_mips_query_stopping · ds	frontiercs2:approximate_mips_query_stopping/ds/run_01M4T7N9B1Y0WSVKKPRWBMWHS2#t0
# Approximate-MIPS Query Stopping `frontier-submit train` returns immediately after acceptance with a verification ID and the authoritative count of remaining train calls. It does not wait for verification. Use `frontier-submit status <verification_id>` for one status read; once verification completes, it includes you...	The task's evaluated harness is different from the shipped main.cpp (which uses Alg0_mfalsh with Qnum=100 and different arguments). Wait — the README says the evaluator invokes `./fg <data-base-path> <ground-truth-path> <policy-resource-path> <suite-id> 150`. But the shipped main.cpp takes `argv[1]` as dataset name and...	approximate_mips_query_stopping · ds	frontiercs2:approximate_mips_query_stopping/ds/run_01M4T7N9B1Y0WSVKKPRWBMWHS2#t1
# Approximate-MIPS Query Stopping `frontier-submit train` returns immediately after acceptance with a verification ID and the authoritative count of remaining train calls. It does not wait for verification. Use `frontier-submit status <verification_id>` for one status read; once verification completes, it includes you...	The Preprocess constructor appends "_new" to the path. So the evaluator passes the data base path directly and Preprocess appends "_new". That matches the README: "data-base-path omits the `_new` suffix, which the loader appends." But the shipped main.cpp is different (it appends ".data"). So indeed the evaluator must...	approximate_mips_query_stopping · ds	frontiercs2:approximate_mips_query_stopping/ds/run_01M4T7N9B1Y0WSVKKPRWBMWHS2#t2
# Approximate-MIPS Query Stopping `frontier-submit train` returns immediately after acceptance with a verification ID and the authoritative count of remaining train calls. It does not wait for verification. Use `frontier-submit status <verification_id>` for one status read; once verification completes, it includes you...	Now let me think about the core scientific problem. We need to write `should_stop` in `src/alg.cpp`. The policy sees SearchState fields per call: - top_scores: pointer to live heap array (unsorted heap representation) - top_score_count: size (number of entries) - kth_score: heap threshold tracked by loop, may be -1 - ...	approximate_mips_query_stopping · ds	frontiercs2:approximate_mips_query_stopping/ds/run_01M4T7N9B1Y0WSVKKPRWBMWHS2#t3
# Approximate-MIPS Query Stopping `frontier-submit train` returns immediately after acceptance with a verification ID and the authoritative count of remaining train calls. It does not wait for verification. Use `frontier-submit status <verification_id>` for one status read; once verification completes, it includes you...	Now I understand the mechanics. Key facts: - `frontier_score = ip1.score` = sum over tables of the bucket's weight contributions: - initial bucket (matching bucket): score = sum over i of weigh[i][0].val where weigh[i][j].val = \|hashval[i*K+j]\| / norm... wait, hashval is divided by norm, so weigh values are \|q·a_j\| /...	approximate_mips_query_stopping · ds	frontiercs2:approximate_mips_query_stopping/ds/run_01M4T7N9B1Y0WSVKKPRWBMWHS2#t4
# Approximate-MIPS Query Stopping `frontier-submit train` returns immediately after acceptance with a verification ID and the authoritative count of remaining train calls. It does not wait for verification. Use `frontier-submit status <verification_id>` for one status read; once verification completes, it includes you...	Now, let's think hard about the algorithm design. Facts about the call sequence in `knnF`: - Initial buckets (matching key in each of the 5 tables) are consumed in the first loop, then the probing sequence loop begins. Each iteration: take next bucket (from precomputed ProbingSequence if not exhausted, else from globa...	approximate_mips_query_stopping · ds	frontiercs2:approximate_mips_query_stopping/ds/run_01M4T7N9B1Y0WSVKKPRWBMWHS2#t5
# Approximate-MIPS Query Stopping `frontier-submit train` returns immediately after acceptance with a verification ID and the authoritative count of remaining train calls. It does not wait for verification. Use `frontier-submit status <verification_id>` for one status read; once verification completes, it includes you...	This confirms the metric computation (upstream version). The evaluator's harness computes recall@50 and kth_ratio similarly: intersection of the first m returned IDs with the exact first m IDs divided by 50; kth ratio = returned inner product at rank m / exact inner product at rank m. Now let's build the lab. First, l...	approximate_mips_query_stopping · ds	frontiercs2:approximate_mips_query_stopping/ds/run_01M4T7N9B1Y0WSVKKPRWBMWHS2#t6
# Approximate-MIPS Query Stopping `frontier-submit train` returns immediately after acceptance with a verification ID and the authoritative count of remaining train calls. It does not wait for verification. Use `frontier-submit status <verification_id>` for one status read; once verification completes, it includes you...	No numpy. I'll write the data generator in C++ or pure Python (slow for 300×54387×192 = 3.1e9 ops in pure Python — way too slow). Better: write a C++ generator that produces the data file and the GT file. Let me write a C++ generator: - Generate N=54387, dim=192 vectors with a chosen distribution (Gaussian, or with no...	approximate_mips_query_stopping · ds	frontiercs2:approximate_mips_query_stopping/ds/run_01M4T7N9B1Y0WSVKKPRWBMWHS2#t7
# Approximate-MIPS Query Stopping `frontier-submit train` returns immediately after acceptance with a verification ID and the authoritative count of remaining train calls. It does not wait for verification. Use `frontier-submit status <verification_id>` for one status read; once verification completes, it includes you...	Now the lab harness main.cpp. It should mimic the evaluator: 5 args, prints FARGO_METRIC lines. Let me write it, and also compute the exact metrics described in the README. Note: the evaluator's exact computation of metrics may differ slightly, but for local comparisons this is fine. Also I need a fake function table...	approximate_mips_query_stopping · ds	frontiercs2:approximate_mips_query_stopping/ds/run_01M4T7N9B1Y0WSVKKPRWBMWHS2#t8
"# Approximate-MIPS Query Stopping\n\n`frontier-submit train` returns immediately after acceptance w(...TRUNCATED)	"Now let me write the lab's alg.cpp with policy variants selectable by macro. But first, let me gene(...TRUNCATED)	approximate_mips_query_stopping · ds	frontiercs2:approximate_mips_query_stopping/ds/run_01M4T7N9B1Y0WSVKKPRWBMWHS2#t9

End of preview. Expand in Data Studio

frontiercs2-t3

T3-shaped retrieval records extracted from FrontierCS-2-runs agent traces.

36,792 records — one per substantive assistant reasoning turn (≥300 chars), across 48 FrontierCS-2 tasks and 16 models.

Format

One JSON object per line, matching the t3rag JSONL ingest schema:

field	content
`question`	the task's `README.md` (the problem the agent was solving), canonicalized per task
`text`	that assistant turn's `reasoning_content` ∪ `content` — the raw reasoning trace
`title`	`<task> · <model>`
`source`	`frontiercs2:<task>/<model>/<run_id>#t<turn>` — full provenance

Provenance

Source: FrontierCS2/FrontierCS-2-runs (CC-BY-4.0), evidence.tar.zst per run
Extraction: last accumulated models/primary/traffic/NNNNN.request.bin per run; task README recovered by correlating tool_calls arguments with tool results across harness formats
Models emitted via OpenAI-style reasoning_content/content fields are included; Claude Code event-stream runs (fable51, opus, sonnet55) and zero-yield runs (astra, muse, sol, terra, wonder) are excluded

Role in a T3 corpus set

This is the raw-traces leg (analogous to raw-chunked): untransformed reasoning text. Struct/reflect/semantic LLM-transformed variants are a separate downstream step.

Per-model record counts

qwen 4,992 · inkling 4,940 · ds 4,763 · glm 4,079 · glmf 3,871 · k3 3,547 · gemini 2,832 · inkling-opencode 2,208 · k27 2,073 · k31 1,974 · minimax 765 · step 450 · inkling-cc 176 · carbon-g 92 · gemini-agy 42 · ds1017 4

Downloads last month: 43