Large-scale agent training needs realistic, tool-using conversation traces rather than synthetic single-turn prompts. This dataset supplies 19,331 multi-turn ChatML traces generated by DeepSeek V4 Pro that include explicit Hermes-style reasoning and extensive tool-call events — a practical resource when you want to teach local models how to act as Hermes agent instances.
What Sets It Apart
- High density of tool interactions: ~138K total tool usages across traces, with detailed counts for core and Hermes tools (e.g., terminal, read_file, web_search, memory). This makes it valuable for training models to manage tool selection and multi-step tool workflows.
- VRAM-tiered variants: four sequence-length / GPU-targeted variants (nano, budget, standard, spark) let practitioners pick datasets matched to LoRA/QLoRA budgets and target token lengths.
- Production-style generation metadata: traces were produced by DeepSeek V4 Pro with 96 parallel workers, a JSON repair filter, and explicit “think” blocks and ChatML formatting — useful for supervised fine-tuning that expects structured agent reasoning.
- Ready-to-use parquet splits: train/valid/test splits and standard parquet/optimized-parquet formats simplify ingestion into Datasets/pandas/polars pipelines.
Who it's for — fit & tradeoffs
Great fit if you want to fine-tune or LoRA-adapt local LLMs to behave as tool-using agents (Hermes-style), particularly when you need realistic multi-turn tool-call sequences and explicit reasoning traces. Also useful for researchers studying tool-usage patterns, SFT workflows, or agent orchestration. Look elsewhere if you need human-authored instruction-following datasets (these traces are machine-generated), extremely large-scale corpora beyond ~20K traces, or multimodal (image/audio) data — this dataset is text/tabular and generated by a specific agent stack, so it reflects that stack's behaviors and biases.
Where it fits
Use this dataset as a mid-sized, engineering-focused resource for LoRA/SFT experiments aimed at agent/tool competency, or as augmentation data when building agent skill managers and tool routers. Treat it as synthetic but structured agent behavior that can accelerate iteration on tool-calling and session-management capabilities.
