CoolFace
Datasetpublic

steviel/qwen3-1.7b-polaris-fp8-rollouts-20260912

Qwen3-1.7B Polaris FP8 rollouts Snapshot of three FP8 rollout datasets taken on 2026-09-12. The model is Qwen3-1.7B-Base. Original Parquet files and attempt, file, and checkpoint-lineage metadata are preserved without rewriting. Configuration Training batches Training responses Validation responses Total bytes maxrl_strict 101 827392 144320 6654288117 maxrl_permissive 107 876544 144320 8168064932 dppo 126 1032192 173184 6733682534 Provenance and… See the full description on the dataset page: https://huggingface.co/datasets/steviel/qwen3-1.7b-polaris-fp8-rollouts-20260912.

sourceHugging Faceupdated 15d agoView on Hugging Face
0likes476downloads
Dataset Card

Qwen3-1.7B Polaris FP8 rollouts

Snapshot of three FP8 rollout datasets taken on 2026-09-12. The model is Qwen3-1.7B-Base. Original Parquet files and attempt, file, and checkpoint-lineage metadata are preserved without rewriting.

ConfigurationTraining batchesTraining responsesValidation responsesTotal bytes
maxrl_strict1018273921443206654288117
maxrl_permissive1078765441443208168064932
dppo12610321921731846733682534

Provenance and interpretation

  • —maxrl_strict: no-baseline MaxRL, strict terminal-box verifier; job 2563052, training steps 1–101.
  • —maxrl_permissive: no-baseline MaxRL, original permissive verifier; job 2556834, training steps 1–107.
  • —dppo: DPPO-TV, original permissive verifier; jobs 2558340 and 2567598, 126 saved training batches through step 111.
  • —Each training batch has 256 prompts × 32 responses (8,192 rows). Validation files contain 28,864 responses each.
  • —DPPO includes overlapping attempts. The original attempt contains superseded steps 65–79; the resumed attempt starts at step 65. These are separate generated samples and are retained here. Use attempt directories/job IDs to construct a coherent training lineage (original through 64, resumed from 65).
  • —Saved rollout steps indicate generated batches, not a guarantee that the corresponding optimizer update or checkpoint completed.
  • —Training and validation use separate configurations because their original schemas differ. The three main configurations expose the train split; append _validation to the configuration name to load the validation split. Validation contains repeated evaluations of the same benchmark prompts across steps; rows are not independent unique questions.
  • —Fields include prompt (input), generated text (output), ground truth (gts), reward (score), step, row index, UID, token IDs, and rollout log probabilities. Strict-verifier telemetry is present in the strict configuration.
  • —Source manifests provide SHA-256 checksums, job IDs, source code commit IDs, and rollout settings. files.json records the exact frozen upload inventory and inventory.json summarizes it.
  • —No new license grant is made by this upload. Underlying source-data terms continue to apply.
python
from datasets import load_dataset
ds = load_dataset("steviel/qwen3-1.7b-polaris-fp8-rollouts-20260912", "maxrl_strict", split="train")