AETHORIA-AI/TR-HASH-Agentic-SFT-32K-210K

TR-HASH Agentic SFT 32K

Balanced instruction and tool-use SFT data for `AETHORIA-AI/TR-HASH-Tokenizer-32K-Agentic`. The canonical repository name is retained, while its contents replace the former tool-heavy 21K laboratory corpus.

Composition

The 9% tool-aware training slice contains tool calls, no-call decisions with tools present, and final answers after tool results. The remaining 91% teaches ordinary instruction following without encouraging tool hallucination.

Every row contains prompt, completion, source lineage, category and token count. Direct answers use <|final_start|>...<|final_end|>. Tool calls use native <|tool_call_start|>...<|tool_call_end|> markers. No target contains a synthetic chain-of-thought or <|think_start|> marker.

Sources and licenses

• —HuggingFaceTB/smol-smoltalk — Apache-2.0.
• —nvidia/When2Call — CC BY 4.0.
• —qgallouedec/tool-calls-mini — Apache-2.0.

This is a mixed-source compilation. Each upstream item remains governed by its source license; users must preserve the applicable attribution and notices. Pinned source revisions, exact counts, token statistics and rejection reasons are recorded in dataset_info.json.

Quality controls

• —official upstream train/test splits are used when available;
• —the small tool-trajectory source is split by stable conversation hash;
• —exact duplicate prompts and cross-split source leakage are rejected;
• —examples over 2,048 tokens are rejected rather than truncated;
• —tool-call JSON and native marker balance are validated;
• —exact PIQA probe exclusion is mandatory for production builds.

Pre-tokenized 2,048-token shards

The native training representation is published under tokenized/tr-hash-agentic-32k-2048/. It contains 150,878,489 train tokens (52,643,063 supervised) and 7,859,965 validation tokens (2,626,627 supervised). Prompt and history labels are masked; only the final assistant target is supervised. No example is truncated during compilation.