Datasets:

hbin0701
/

opsd-probe-seed

pid int64 61 89	ridx int64 0 63	frac float64 0.25 0.75	src_correct bool 2 classes	n_prefix int64 4.22k 29.2k	n_think_total int64 5.63k 38.9k	completions int64 4 4	prefix_source stringclasses 1 value	base_correct int64 0 4	champion_correct int64 0 4
80	40	0.75	false	29,184	38,912	4	base	0	2
84	39	0.75	false	29,184	38,912	4	base	0	0
80	31	0.75	false	28,614	38,152	4	base	0	0
80	4	0.75	true	24,973	33,297	4	base	3	1
76	60	0.75	false	23,602	31,469	4	base	0	0
80	16	0.75	false	21,712	28,949	4	base	0	0
76	19	0.75	false	21,652	28,870	4	base	1	0
80	10	0.75	false	21,392	28,522	4	base	1	1
88	40	0.75	false	20,608	27,477	4	base	0	0
88	19	0.75	false	20,289	27,052	4	base	0	0
88	44	0.75	false	19,883	26,511	4	base	0	0
76	0	0.75	false	19,554	26,072	4	base	0	0
80	40	0.5	false	19,456	38,912	4	base	1	1
84	39	0.5	false	19,456	38,912	4	base	0	0
76	58	0.75	false	19,246	25,661	4	base	0	0
80	31	0.5	false	19,076	38,152	4	base	0	0
88	30	0.75	false	18,656	24,874	4	base	0	0
88	36	0.75	false	18,391	24,521	4	base	0	0
88	57	0.75	false	18,328	24,437	4	base	0	0
88	0	0.75	false	18,094	24,125	4	base	0	0
68	50	0.75	false	17,502	23,336	4	base	0	1
68	30	0.75	false	17,417	23,223	4	base	1	0
76	8	0.75	true	17,248	22,997	4	base	0	0
76	9	0.75	true	16,851	22,468	4	base	4	3
80	4	0.5	true	16,648	33,297	4	base	1	3
68	38	0.75	true	16,505	22,007	4	base	4	2
76	60	0.5	false	15,734	31,469	4	base	0	1
68	16	0.75	false	15,167	20,223	4	base	1	0
76	56	0.75	true	14,727	19,636	4	base	3	3
80	16	0.5	false	14,474	28,949	4	base	0	0
76	19	0.5	false	14,435	28,870	4	base	1	0
80	10	0.5	false	14,261	28,522	4	base	2	2
80	63	0.75	true	14,120	18,826	4	base	4	4
88	40	0.5	false	13,738	27,477	4	base	0	0
88	19	0.5	false	13,526	27,052	4	base	0	0
88	44	0.5	false	13,256	26,511	4	base	0	0
76	0	0.5	false	13,036	26,072	4	base	1	3
76	58	0.5	false	12,830	25,661	4	base	1	0
88	6	0.75	true	12,487	16,649	4	base	3	3
88	30	0.5	false	12,437	24,874	4	base	0	0
88	36	0.5	false	12,260	24,521	4	base	0	0
88	57	0.5	false	12,218	24,437	4	base	0	0
88	0	0.5	false	12,062	24,125	4	base	0	0
72	0	0.75	false	11,849	15,799	4	base	2	1
68	50	0.5	false	11,668	23,336	4	base	0	1
68	30	0.5	false	11,612	23,223	4	base	0	2
68	56	0.75	false	11,480	15,307	4	base	4	4
76	8	0.5	true	11,498	22,997	4	base	0	0
76	9	0.5	true	11,234	22,468	4	base	0	0
80	43	0.75	true	11,065	14,753	4	base	4	4
68	38	0.5	true	11,004	22,007	4	base	0	1
72	15	0.75	true	10,734	14,312	4	base	4	4
68	16	0.5	false	10,112	20,223	4	base	2	1
80	40	0.25	false	9,728	38,912	4	base	1	3
84	39	0.25	false	9,728	38,912	4	base	0	1
76	56	0.5	true	9,818	19,636	4	base	4	3
80	31	0.25	false	9,538	38,152	4	base	1	1
80	27	0.75	true	9,526	12,701	4	base	4	4
80	63	0.5	true	9,413	18,826	4	base	1	4
64	43	0.75	true	8,857	11,809	4	base	4	4
64	24	0.75	true	8,680	11,573	4	base	4	4
88	6	0.5	true	8,324	16,649	4	base	3	2
64	38	0.75	false	8,643	11,524	4	base	0	1
80	4	0.25	true	8,324	33,297	4	base	2	1
64	45	0.75	false	8,354	11,139	4	base	0	0
64	34	0.75	true	8,154	10,872	4	base	4	4
64	19	0.75	false	7,925	10,567	4	base	0	1
72	0	0.5	false	7,900	15,799	4	base	3	3
76	60	0.25	false	7,867	31,469	4	base	2	1
64	9	0.75	true	7,721	10,295	4	base	4	4
84	40	0.75	false	7,610	10,146	4	base	4	4
68	56	0.5	false	7,654	15,307	4	base	4	3
80	43	0.5	true	7,376	14,753	4	base	4	4
80	16	0.25	false	7,237	28,949	4	base	1	1
80	10	0.25	false	7,130	28,522	4	base	2	0
76	19	0.25	false	7,218	28,870	4	base	1	2
88	40	0.25	false	6,869	27,477	4	base	0	0
72	15	0.5	true	7,156	14,312	4	base	3	2
88	19	0.25	false	6,763	27,052	4	base	0	0
88	44	0.25	false	6,628	26,511	4	base	0	0
64	14	0.75	false	6,680	8,906	4	base	0	1
88	30	0.25	false	6,218	24,874	4	base	0	0
76	0	0.25	false	6,518	26,072	4	base	3	3
88	36	0.25	false	6,130	24,521	4	base	0	0
88	57	0.25	false	6,109	24,437	4	base	0	0
80	27	0.5	true	6,350	12,701	4	base	4	4
76	58	0.25	false	6,415	25,661	4	base	1	0
88	0	0.25	false	6,031	24,125	4	base	0	0
64	43	0.5	true	5,904	11,809	4	base	3	4
68	50	0.25	false	5,834	23,336	4	base	3	2
68	30	0.25	false	5,806	23,223	4	base	1	3
64	24	0.5	true	5,786	11,573	4	base	2	2
64	38	0.5	false	5,762	11,524	4	base	1	2
76	8	0.25	true	5,749	22,997	4	base	2	2
76	9	0.25	true	5,617	22,468	4	base	0	0
64	45	0.5	false	5,570	11,139	4	base	1	0
68	38	0.25	true	5,502	22,007	4	base	4	3
68	36	0.75	true	5,428	7,237	4	base	4	4
64	34	0.5	true	5,436	10,872	4	base	1	3
68	54	0.75	true	5,330	7,106	4	base	4	4

End of preview. Expand in Data Studio

OPSD prefix-continuation probe — seed data

Everything needed to reproduce the prefix-continuation probe for OPSD (on-policy self-distillation) on a fresh GPU box, except the base model (Qwen/Qwen3-1.7B, pulled from the Hub at setup) and the code repo (hbin0701/OPSD).

These artifacts live outside git because the training/eval output directory is .gitignored.

What the probe answers

Fitting p' = p + λ·(1[mode correct] − p) + γ against a properly sampled 64-shot baseline puts λ (the pull toward the model's own modal answer) at 0.128, with a 95% interval spanning zero, and γ — a flat gain collected everywhere — at +5.83 pp, 68% of the total +8.59 pp gain. So most of what OPSD buys is not mode-sharpening, and every measurement in the project scores whole attempts, which cannot separate "picks a better answer" from "executes an attempt better".

The probe forces that separation: stored rollouts are truncated mid-thinking, and both policies complete the same prefixes inside one vLLM engine (base = no adapter, champion = LoRA request), so batching, sampling and context are identical by construction.

	base completes	champion completes
base prefixes	control	execution effect
champion prefixes	prefix effect	both

Row effect = execution quality. Column effect = trajectory quality.

Contents (`opsd_probe_seed.tar.gz`, 148 MB → 350 MB unpacked)

path	what
`exp/runs/think_ablation/dpeval/base_at64/shard*.json`	untrained Qwen3-1.7B on AIME24, 30 problems × 64 samples, full generation text. avg 48.44 / pass 83.33 / maj 70.00
`exp/runs/think_ablation/dpeval/qwen31b_incorrect_only_noopen_100steps_ckpt100_at64/shard*.json`	the OPSD champion, same protocol. avg 57.03 / pass 83.33 / maj 70.00
`exp/runs/incorrect_only/.../checkpoint-100/adapter_{config.json,model.safetensors}`	the champion LoRA (r=64). Inference files only — the 1.5 GB DeepSpeed optimiser state is deliberately excluded
`exp/src/tokencap/rollouts.json`	300 cached on-policy training rollouts with prompts and token ids, for the clamp/agreement-mass diagnostic
`exp/src/prefix_probe/s_spread.json`	already-computed agreement-mass distribution (300 rollouts, 378,207 positions)

Eval protocol for both shard sets: AIME24, thinking mode on, max_tokens 38912, T=1.0, top_p 0.95, tp=1, 4-way data-parallel, graded with math_verify 0.8.0.

Use

This dataset is data only — the code is not here. The runner lives in the project repo at exp/src/prefix_probe/bootstrap.sh (branch reports/experimental-record); it provisions a bare GPU box, pins the dependency set, downloads the base model, pulls this tarball from the Hub automatically, runs a preflight and launches.

bash bootstrap.sh --run              # probe
bash bootstrap.sh --with-training    # + the training stack

To use the data on its own, the tarball unpacks to repo-relative paths:

wget https://huggingface.co/datasets/hbin0701/opsd-probe-seed/resolve/main/opsd_probe_seed.tar.gz
tar xzf opsd_probe_seed.tar.gz -C <repo-root>

Requires driver ≥ 525. The probe is ~4–6 h on four A100-80GB.

A note on the agreement mass

s_spread.json records the per-position distribution of S = Σ_v 1[e_v < c]·q_v, the quantity that is simultaneously the step size and the support of the target in g = S·(p − q̃). It is not a gentle reweighting: at ≥75% of positions S = 1.0 exactly (nothing clamped), while the bottom percentile falls to 0.055. The clamp is inactive most of the time and closes hard at the opening (mean S 0.831 for tokens 1–64, relaxing to 0.938 past 1k).

Downloads last month: 43

OPSD prefix-continuation probe — seed data

What the probe answers

Contents (opsd_probe_seed.tar.gz, 148 MB → 350 MB unpacked)

Use

A note on the agreement mass

Contents (`opsd_probe_seed.tar.gz`, 148 MB → 350 MB unpacked)