Datasetpublic

armand0e/qwen37-pi-qwen36-27b-topk40-logprobs

Qwen3.7 Pi Trace Top-40 Teacher Logprobs Offline top-40 teacher logprobs for cumulative assistant-turn rows from armand0e/qwen3.7-max-split-formatted. These files are intended to be loaded with snapshot_download, not datasets.load_dataset. Contents manifest.json: shard metadata and filtering counts shard-*.pt: tokenized examples with labels, target positions, top-k token ids, and top-k teacher logprobs chat_template.jinja: the exact chat template used for… See the full description on the dataset page: https://huggingface.co/datasets/armand0e/qwen37-pi-qwen36-27b-topk40-logprobs.

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

0likes148downloads

Dataset Card

Qwen3.7 Pi Trace Top-40 Teacher Logprobs

Offline top-40 teacher logprobs for cumulative assistant-turn rows from armand0e/qwen3.7-max-split-formatted.

These files are intended to be loaded with snapshot_download, not datasets.load_dataset.

—manifest.json: shard metadata and filtering counts
—shard-*.pt: tokenized examples with labels, target positions, top-k token ids, and top-k teacher logprobs
—chat_template.jinja: the exact chat template used for tokenization

Summary

—Source rows: armand0e/qwen3.7-max-split-formatted/default:train
—Teacher: Qwen/Qwen3.6-27B
—Student tokenizer: Qwen/Qwen3.5-9B
—Kept rows: 699 of 1,485
—Examples: 699
—Target positions: 795,348
—Max length: 32,768
—Max target tokens: 8192
—Value type: full-vocabulary logprobs for the selected top-k tokens

The training loss should apply CE to recorded labels and KL against the stored teacher top-k distribution only at target_positions.