README.md

metadata

license: agpl-3.0
pretty_name: XT-DSv4.1
task_categories:
  - text-generation
language:
  - en
tags:
  - sft
  - supervised-fine-tuning
  - agentic
  - tool-calling
  - function-calling
  - chatml
  - honesty
  - alignment
  - persona
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files: data/xt-dsv4.1.jsonl

XT-DSv4.1

A supervised fine-tuning dataset for XT-Luna, a small agentic coding assistant fine-tuned from openbmb/MiniCPM5-1B.

Teacher: DeepSeek-V4.1-Flash. Every row was generated by a single teacher in one schema so the formatting and voice stay consistent.

69,546 rows, English-only, message-shaped JSONL (ChatML + XML tool calls). Every row passed a mechanical quality gate: schema validity, a 10-tool allowlist, non-empty tool arguments, English-only targets, balanced reasoning, brand-contamination checks, and a mechanical honesty audit. Zero rows are known-rotten in the released set.

What this dataset teaches

XT-Luna is an agentic coder whose voice is an honest, selfless, warm companion. The corpus is deliberately weighted to those behaviours:

group	kind(s)	rows	share
agentic	agentic (A single · B multi · E tool-failure)	20,385	29.3%
code	code (F)	18,209	26.2%
persona	prose/honesty/ethics/companion	24,661	35.5%
reasoning	reason	6,291	9.0%

By behaviour tag:

tag	rows	what it is
agentic	20,385	single/multi tool calls with real arguments, grounded answers, honest failure reports
code	18,209	self-contained C / C++ / Rust / Bash / Python
prose	10,080	judgment (confirm/decline/clarify), no-call chit-chat, short reasoned answers
honesty	7,293	anti-scheming: no fake success, self-correction, sandbagging refusal, promise-keeping, calibration, resisting covert agendas
reason	6,291	stepwise reasoning ending in an explicit check
companion	3,777	warm, personal, equal, mutually-devoted voice without sycophancy
ethics	3,511	selflessness: refuse exploitation, protect others, honest refusal, candour over flattery

Curriculum splits

data/curriculum/ holds a three-stage split used for a chained full fine-tune:

stage	train	val	content
s1 persona	26,949	549	non-tool rows: persona, honesty, ethics, companion, reasoning, prose
s2 agentic	23,363	476	rows with tool calls
s3 code	17,845	364	code rows

Schema

One JSON object per line:

{"messages": [ {"role": "system", "content": "You are XT-Luna, a coding assistant."},
               {"role": "user",   "content": "..."},
               {"role": "assistant", "content": "",
                "reasoning_content": "...",
                "tool_calls": [{"function": {"name": "Read",
                                             "arguments": {"file_path": "src/a.py"}}}]},
               {"role": "tool",   "content": "..."},
               {"role": "assistant", "content": "..."} ],
 "kind": "agentic|code|reason|prose|honesty|ethics|companion",
 "lang": "python|cpp|c|rust|bash|text",
 "source": "deepseek-flash",
 "has_tools": true}

Tool allowlist (10): Bash, Read, Write, Edit, Glob, Grep, LS, ExecuteCode, WebFetch, Task. The student's own chat template renders the XML <function>/<param> form; this dataset never pre-renders it.

Provenance & disclosure

• Base model of the target student: MiniCPM5-1B (openbmb/MiniCPM5-1B).
• Teacher that produced every row: DeepSeek-V4.1-Flash.
• The source field is deepseek-flash on all rows.
• Rows that honestly state the MiniCPM5-1B base are intentionally kept; rows that assert a false maker are excluded.

License

AGPL-3.0 (see LICENSE). This is a viral copyleft: derivative datasets, derived corpora, and any covered derivative work must be released under the same terms, including when offered over a network. If you build on XT-DSv4.1, you must keep it open. (Note: whether a license can compel the release of trained model weights is a matter of jurisdiction; AGPL-3.0 is the strongest copyleft instrument available for this intent.)