rasulkhanbayov/truncation-sampler-iclr

Decode logs for "Truncation Samplers Are Nested: Comparing Decoding Rules Is a Measurement Problem"

Per-step decoding logs behind the paper's certificates. Code, configs, results and instructions are in the GitHub repository; use its scripts/hf_download.sh to fetch these files into the paths the code expects.

Contents

Path	Files	What
logs/*.npz	200	One decoding run each: grid walks (*_gwedge_s<seed>.npz), pairwise walks under a pair's own A∧B, ρ-sweep and horizon runs, and a calibration corpus
e02_cache/*.npz	7	Per-configuration cutoff arrays (K for the grid's narrowest and widest rules), restored to results/.e02_cache/
data/MANIFEST.tsv	1	path, bytes and sha256 of every file above, at its restored path
data/LOG_SCOPES.tsv	1	two symlink sets the configs read (recreated by the download script)

About 31 GB in total.

Format

Each logs/*.npz holds, for every decoding step, the top-2048 next-token probabilities in sorted order (probs), the sorted index of the sampled token (token_index), a flag for steps whose cutoff needed the untracked tail (trunc_tail), per-prompt offsets (prompt_ptr), realized lengths, EOS flags, prompt ids, and a JSON meta string recording the model, pinned revision, task, the rule walked under, seed, temperature, dtype, batch size and engine. Read them with code/sepcert/io.py (load_logs).

Models and tasks

Qwen3-0.6B/1.7B/4B/8B, Gemma-3-4B-it, gpt-oss-20b and DeepSeek-R1-Distill-Qwen-1.5B (revisions in configs/models.yaml), on GSM8K, MATH-500, GPQA-Diamond, AIME 2025 and MT-Bench prompts. Batch size 1, bfloat16, temperature 1, Hugging Face Transformers.

Licence

The logs are released under MIT (to be confirmed by the authors). They contain model output probabilities and token indices, not dataset text; the prompts themselves come from the datasets' own sources under their own licences (see configs/tasks.yaml).