Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

PreG per-token signals (math)

Per-token internal signals of every reasoning path in rjie5656/TTS-512Paths, as used for the paper Early Signals, Strong Decisions: Prefix-Guided Sampling for Parallel Test-Time Scaling (arXiv:2611.14965, code: https://github.com/renjie3/PreG). Most users only need the path-level scores in TTS-512Paths/scores/; these files are for token-level analyses.

Layout: <model>/<dataset>/<shard>.pt, models qwen3-8b (Qwen/Qwen3-8B) and llama-nemotron-4b (nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1), datasets AIME25, BRUMO25, HMMT25-Feb, CMIMC25. Concatenate the shards of a dataset in the order of their start index (the first number in the file name). GPQA-Diamond is not included.

Known overlap: for qwen3-8b/hmmt, problem 29 appears both in ..._7680_ckpt.pt and in ..._14336_1.pt; skip problem 29 of ..._14336_1.pt.

Each file is a list with one dict per path (load with torch.load(path, weights_only=False); these are pickles, only load files you trust):

key content
problem_id, sample_id 1-based problem index (line in TTS-512Paths/data/<model>/<dataset>.jsonl), path index
prompt, completion, prediction, answer user prompt, generated text, extracted and gold answer
full_ids token ids of chat prompt + completion, [1, L]
first_pred_pos, last_pred_pos logit positions predicting the first / last completion token
deepconf, self_certainty, entropy, gini, margin per completion token, [T], bf16, temperature 1.1
target_prob probability of the generated token under the logit lens of every layer, [num_layers + 1, T]; the last row is the model output
coe, c Chain-of-Embedding feature (first 300 tokens) and truncation point, used by the ST-BoN baseline

The path-level scores of the paper are computed from these with preg/scores.py in the code release (tools/pt_to_scores.py).

Downloads last month
47

Paper for rjie5656/preg_pt