Dogacel/nemotron-post-training-v2-qwen-3.5-0.8b-regen
Dataset Card for Nemotron Post Training v2 Qwen 3.5 0.8B Regen Regenerated responses from nvidia/Nemotron-Post-Training-Dataset-v2 dataset using Qwen3.5 0.8B model. Parameter Value Max Tokens 4096 Temperature 1.0 Top-k 20 Top-p 0.95 Repetition Penalty 1.5 Dataset consists only the english samples from the Nemotron Post Training Dataset. 85% of the chat prompts have reasoning enabled, every other category has reasoning disabled. Category Value… See the full description on the dataset page: https://huggingface.co/datasets/Dogacel/nemotron-post-training-v2-qwen-3.5-0.8b-regen.
Dataset Card for Nemotron Post Training v2 Qwen 3.5 0.8B Regen
Regenerated responses from nvidia/Nemotron-Post-Training-Dataset-v2 dataset using Qwen3.5 0.8B model.
Dataset consists only the english samples from the Nemotron Post Training Dataset. 85% of the chat prompts have reasoning enabled, every other category has reasoning disabled.
Total dataset excludes 1341 samples from the configuration above, which exceeded the maximum pre-fill length of 16K tokens.
Dataset Creation
Regenerated using sglang version 0.5.10.post1 on 8xH200 cluster in 10 hours.
python -m sglang.launch_server \
--model "Qwen/Qwen3.5-0.8B" \
--reasoning-parser "qwen3" \
--port "30000" \
--dp-size "8" --tp-size "1" \
--dtype "bfloat16" \
--mem-fraction-static "0.9" \
--cuda-graph-max-bs "1024" \
--max-running-requests "1024" \
--mamba-full-memory-ratio "9.0" \
--context-length "16384" \
--max-prefill-tokens "8192" \
--log-level "info" \
--log-level-http "warning" \
--decode-log-interval "200" \
--trust-remote-code \
--load-balance-method "round_robin"Samples were regenerated using SpecForge format.
python "$ROOT_DIR/scripts/regenerate_train_data.py" \
--model "Qwen/Qwen3.5-0.8B" \
--concurrency "9216" \
--max-tokens "4096" \
--temperature "1.0" \
--top-p "0.95" \
--top-k "20" \
--repetition-penalty "1.5" \
--server-address "localhost:30000" \
--input-file-path "nemotron_post_training_v2_prompts.jsonl" \
--output-file-path "nemotron_post_training_v2_qwen35_0.8B.jsonl" \
--resumeSource Data
This dataset is derived from Nemotron-Post-Training-Dataset-v2 by NVIDIA. Prompts retain their original licenses (CC-BY-4.0, ODC-BY for Wildchat-sourced prompts, CC-BY-SA for StackOverflow-sourced prompts) — see the license field on each sample.
Acknowledgements
We would like to thank Fal.ai for providing the necessary compute resources.