Author: hf-mirror · License: other (synthetic research corpus; verify DeepSeek API terms for commercial use) · Format: parquet
1 downloads · 0 stars · 0 mirrors · 22,616,706 · 1 file
19,072 verifier-passed agentic traces from DeepSeek-V4-Pro 0813 across 13 verifiable task families (SFT + GRPO/RLVR ready).
Tags: agentic · tool-calling · distilled · verifiable · grpo · deepseek-v4 · nemotron-ready
Verifiable-first agentic training corpus: 19,072 training traces (plus 2,135 held-out eval rows) from DeepSeek-V4-Pro 0813 in thinking mode, across 13 verifiable task families — tool-call schema adherence with real sandbox execution, verified constraint instructions, structured outputs, stateful dialogue, and more. Every row passed a deterministic programmatic verifier and re-verified on replay; 27,541 real tool-call steps, parallel-call and error-recovery lanes included; aligned with the NVIDIA Nemotron 3 Ultra recipe and NeMo Gym JSONL contract. License tag is 'other' (synthetic research corpus) — check DeepSeek's API terms before commercial training use.
Direct download: https://huggingface.co/datasets/r0b0tlab/deepseek-v4-pro-0813-agentic/resolve/main/canonical/train-00000-of-00001.parquet
Or via the API: curl -L -o file "https://huggingface.co/datasets/r0b0tlab/deepseek-v4-pro-0813-agentic/resolve/main/canonical/train-00000-of-00001.parquet"
Added 2026-09-17. Platform: Distillio — the open source training data collective.