Datasetpublic

atenareply/asterion-agentic-sft

Asterion Agentic SFT (v2, portable) What's inside — 19,828 chat conversations for agentic SFT, sized for a 12B model: grounded analysis QA on Asterion Space Operations docs (track A, 2,997, stratified over 7 doc_type x 12 topic), multi-turn tool-calling trajectories for telemetry alert triage (track B, 6,000, difficulty EXACTLY 20/30/30/20 = 1,200/1,800/1,800/1,200), raw-log multi-QA (track C, 917, four qa_type: pointwise/aggregation/selection/reasoning), plus an equal 9,914-row… See the full description on the dataset page: https://huggingface.co/datasets/atenareply/asterion-agentic-sft.

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes29downloads

Dataset Card

Asterion Agentic SFT (v2, portable)

What's inside — 19,828 chat conversations for agentic SFT, sized for a 12B model: grounded analysis QA on Asterion Space Operations docs (track A, 2,997, stratified over 7 doctype x 12 topic), multi-turn tool-calling trajectories for telemetry alert triage (track B, 6,000, difficulty EXACTLY 20/30/30/20 = 1,200/1,800/1,800/1,200), raw-log multi-QA (track C, 917, four qatype: pointwise/aggregation/selection/reasoning), plus an equal 9,914-row general blend from smol-smoltalk.

Where it comes from — asterion-cpt-corpus (pinned revision e0911591) + mars-logs-curated + mars-express-telemetry-logs supply the domain half; the teacher (Scaleway, gpt-oss-120b, temp 0.4) wrote only questions/answers/reasoning, while ALL tool calls and tool results were produced by code (grounded-by-construction).

How it was built / modified — v2 fixes over noval-corp-agentic-sft: (a) raw-vs-clean tool output is an INDEPENDENT per-task coin flip (rate 0.5) — kills the v1 'raw => genuine' confound; (b) track C generates K=4 typed QA per log excerpt; (c) teacher reasoning is stored in the STRUCTURED reasoning field (no inline <think> markup) so each model's training template renders it natively. Stored model-agnostically as one JSON string per row (messages_json) with OpenAI-style structured tool_calls — the same data trains LFM2.5-1.2B AND Gemma-4-12B.

Overview

—Role: SFT dataset — Asterion round (shared: LFM2.5-1.2B and Gemma-4-12B)
—Lineage: asterion-cpt-corpus (pinned) + mars-logs-curated + mars-express-telemetry-logs + smol-smoltalk → buildagenticsft_v2.py (teacher = Scaleway) → this → lfm2.5-1.2b-asterion-agentic (SFT) and, next round, gemma-4-12b-asterion-agentic

Schema

Column	Meaning
`messages_json`	the whole conversation, JSON-serialized (system/user/assistant with structured tool_calls/tool turns; assistant may carry `reasoning`)
`track`	`A` analysis QA	`B` tool trajectory	`C` raw-log multi-QA	`general`
`difficulty`	0 (non-agentic) or 1-4 curriculum tier (Track B)
`qa_type`	track C only: pointwise	aggregation	selection	reasoning

Splits

Split	Rows	Note
train	18,837	9,914 domain + ~8,923 general blend
validation	991	domain held-out (never general)

Used by

—trainagenticv2.py → atenareply/lfm2.5-1.2b-asterion-agentic
—traingemmaagentic.py (future Gemma-4-12B round, same file re-used)
—evalagenticv2.py / grpotaskslfm2_asterion.py share the task semantics

Limitations

—Track C yielded 917 of the 3,000-record target (multi-QA survival is lower under the grounding filters); track A/B hit target (2,997/6,000).
—The genuine-anomaly class is largely synthetic; telemetry values come from Mars Express, not from a real Asterion fleet.
—Do NOT load messages_json as a struct column: HF datasets would union tool-argument keys across rows and inject nulls — parse the JSON string.

Part of the **noval-corp** fictional-domain training family (Orbital Mining Corporation / Mars Express; Asterion round for Gemma-4). Card generated by `noval-corp/scripts/gendatasetcards.py` — see `noval-corp/NOTE.md` for the full project log.