qualia1-sft

Per-document SFT sessions: an assistant that knows one document and uses it everywhere.

qualia1-sft teaches a model that carries a docs1 document (in its weights, through a hypernetwork, or in its context) to act as a general assistant and agent that uses what it knows: recalling it word for word, answering, reasoning, summarizing, writing code, calling tools, and saying "I don't know" when the document does not say. Each session is a self-distillation target: the answers were written by MiniCPM5-2B with the document in its context.

Sessions	1,977,861
Documents (docs1)	99,998 (about 20 sessions each)
Generated tokens (thinking + answers + tool calls)	3.73B
User turns with thinking on	90%
Families / archetypes	14 / 797 distinct archetype labels
docs1 revision	8c384a2237f4757dad0354e1169283a441dc8ad4 (sha256 of every document verified)
Built	2026-10-15

Usage

Rows reference documents by doc_id and pin docs1_revision and doc_sha256; the document text is not copied. Join once before training:

import json

from datasets import load_dataset

sessions = load_dataset("alpha-qualia/qualia1-sft", split="train")
docs = load_dataset("alpha-qualia/docs1", split="train", revision="8c384a2237f4757dad0354e1169283a441dc8ad4",
                    data_files="data/b0[0-2]_*/*.parquet")  # all qualia1-sft documents are 1,024-4,095 tokens
text = {d["id"]: d["content"] for d in docs}
row = sessions[0]
document = text[row["doc_id"]]           # what the model should know
messages = json.loads(row["messages"])   # what it should say (student view: no document in the prompt)
tools = json.loads(row["tools"]) if row["tools"] else None

messages is OpenAI-style chat (system, user, assistant with reasoning_content and tool_calls, tool) and renders with the MiniCPM5 chat template (<think>, <function name=…><param …>, <tool_response>).

How it was made

1. Documents. 99,998 docs1 documents of 1,024–4,095 tokens, equal per length bucket, weighted by source category (code, agent files, docs and textbooks 1.5×; general, science, transcripts 1.0×; literature and law 0.5×).
2. Skeletons (skeletons config). Code sampled ~20 session slots per document (family, turns, thinking switch, user voice, how the user mentions the document, length and format asks, system prompt kind, difficulty, evidence passages). gpt-6-luna (medium effort, Batch API) read the document and wrote, for each slot, the system prompt, every user turn, and for agentic sessions 3–8 invented tools and their mock results. Follow-up turns never depend on the answer.
3. Bodies (this config). MiniCPM5-2B with the document in its system prompt answered every turn: its own thinking (when the turn's switch is on), answers and tool calls. Agentic sessions ran in a mock environment: a call returns that tool's next mock result (arguments not matched); an unknown or exhausted tool returns {"error": "not available"}; at most 8 tool rounds per turn. vLLM 0.29, temperature 1.0, top-p 0.95, FlashInfer with an fp8 KV cache, 8× RTX PRO 6000.
4. No filters. Every answer is kept, right or wrong; status records how the session ended.

The teacher prompt began with You have the following document in memory. Use it when it is relevant. followed by the document, then the session's own system prompt. The stored messages carry only the session's system prompt.

Families

Family	Sessions	% sessions	% tokens	Median tokens	p90 tokens	Length-cut
reconstruct	282,580	14.3%	24.7%	2,790	6,514	1.8%
recite	184,764	9.3%	4.3%	533	1,841	0.4%
agentic	168,288	8.5%	5.6%	798	2,677	0.2%
reasoning	166,552	8.4%	8.5%	1,232	4,323	0.8%
indirect	150,128	7.6%	9.5%	1,521	5,392	0.9%
fact	150,087	7.6%	3.8%	511	2,161	0.3%
code	150,042	7.6%	12.6%	1,955	7,820	2.9%
calibration	149,298	7.5%	4.3%	675	2,297	0.4%
summarize	138,604	7.0%	6.4%	1,078	3,904	0.5%
transform	127,010	6.4%	6.3%	1,109	4,279	0.7%
unrelated	94,971	4.8%	3.7%	615	3,817	1.3%
create	86,399	4.4%	4.5%	1,138	4,569	0.6%
dialogue	69,217	3.5%	3.5%	1,236	4,085	0.6%
meta	59,921	3.0%	2.4%	956	3,347	0.5%

Reconstruction (whole document and targeted recitation) is native: many differently worded requests written by Luna, answered from the document in context, not a fixed template.

Archetypes

Luna picked the best-fitting archetype of the sampled family for each document (labels with at least 98 sessions; the full list is in stats.json).

Family	Archetype	Sessions
reconstruct	the whole document retold in their own words in the original order	116,339
reconstruct	the whole document as close to the original as possible	91,273
reconstruct	the whole document as faithfully as possible but in another format (for example notes or a list)	24,515
reconstruct	the whole document lightly rephrased	21,686
reconstruct	everything it says in order with nothing left out	17,212
reconstruct	the whole document word for word	10,845
reconstruct	the whole document faithfully as possible but in another format (for example notes or a list)	552
recite	quote the passage that says a given thing	111,142
recite	recite the section under a given heading	27,229
recite	recite a list or table from it	26,274
recite	continue the text from a given phrase	10,572
recite	give its opening or closing lines	4,356
recite	recite a section under a given heading	3,378
recite	fill in a blank in one of its sentences	1,366
recite	recite a passage that says a given thing	148
fact	an exact number date name or identifier	71,279
fact	a filtered list of all items with some property	33,648
fact	who what or where	23,722
fact	the definition of a term as it defines it	11,288
fact	check whether it makes a given claim (sometimes it does not)	8,174
fact	look up an entity from an informal description	579
fact	a detail it does not contain	256
fact	explain why or what caused something	105
reasoning	join two or three facts	48,234
reasoning	explain why or what caused something	32,147
reasoning	compare two entities or options	30,733
reasoning	apply its rule or method to a new case	21,199
reasoning	compute something from its numbers	16,365
reasoning	spot a tension or inconsistency in it	7,137
reasoning	what-if with one fact changed	5,322
reasoning	order events on a timeline	5,263
summarize	key points	41,738
summarize	only about one subtopic	39,837
summarize	practical takeaways	20,717
summarize	one-line summary	15,136
summarize	section by section	13,980
summarize	abstract-style paragraph	2,972
summarize	outline	1,621
summarize	for a child	1,404
summarize	for an expert	1,034
transform	paraphrase a passage	41,688
transform	turn part of it into a table	35,586
transform	FAQ	11,091
transform	rewrite a passage in another style or register	9,370
transform	turn part of it into JSON or YAML	8,753
transform	timeline	8,724
transform	flashcards	7,688
transform	quiz questions with answers	2,317
transform	translate a passage	567
transform	turn part of it into a timeline	287
transform	turn part of it into bullets	222
transform	turn part of it into numbered steps	165
transform	turn part of it into a numbered list	123
indirect	a plan or how-to for the user's situation	64,720
indirect	advice for a decision the user faces	43,150
indirect	the user's own problem where its knowledge helps	23,414
indirect	a recommendation	11,333
indirect	sanity-check the user's claim or plan	7,316
create	a social post or announcement	24,259
create	a talk or presentation outline	18,864
create	an email or message built on it	16,209
create	brainstorm follow-up ideas	12,140
create	a critique or review of it	5,744
create	a short story or poem inspired by it	4,728
create	a lesson plan	4,396
code	write code using an API or procedure it describes	80,137
code	a script that parses or computes over its data	22,889
code	a config file or shell command following it	15,017
code	write tests for its code	12,956
code	fix the user's code according to it	12,683
code	explain code from it	5,241
code	port its code to another language	582
code	a general coding task	155
code	a config file or shell command following it describes	111
agentic	memory instead of a tool: tools are offered but the answer is already known	83,215
agentic	act for the user: send, schedule or file something built from it	26,599
agentic	memory fills tool arguments: names, dates, paths or values from it go into a call	24,093
agentic	act for the user	8,515
agentic	multi-step exploration: search, read, call again	8,035
agentic	memory instead of a tool: the answer is already known	3,617
agentic	act for the user: send, schedule or file something built on it	3,024
agentic	a tool result conflicts with it: the result is fresher and disagrees	2,905
agentic	a tool fails or returns nothing: fall back to memory	2,458
agentic	memory instead of a tool: the answer is in the document; tools are available but unnecessary	2,283
agentic	memory instead of a tool: the answer is in the document	857
agentic	memory fills tool arguments: names, dates or values from it go into a call	548
agentic	act for the user: a tool that sends, schedules or files something built from it	481
agentic	act for the user: a tool that sends, schedules or files, returning a confirmation	477
agentic	memory fills tool arguments	415
agentic	a tool result conflicts with it: a result is fresher and disagrees	120
dialogue	drill down with follow-ups	44,452
dialogue	the user states something wrong about it	15,115
dialogue	the user challenges with are-you-sure	4,087
dialogue	the user corrects their own earlier request	4,010
dialogue	revise the answer (shorter or another angle)	1,410
calibration	a detail it does not contain	94,738
calibration	a false premise it contradicts or never states	44,547
calibration	a plainly unknowable question	5,582
calibration	whether it is still true today	2,391
calibration	a question only partly answerable from it	1,640
calibration	check whether it makes a given claim (sometimes it does not)	271
unrelated	a math problem	35,967
unrelated	a general coding task	19,257
unrelated	a logic puzzle	16,293
unrelated	a writing or editing task with the text supplied	16,149
unrelated	a general knowledge question	4,896
unrelated	a word that also appears in it but with another meaning	1,768
unrelated	casual chat	574
meta	how it is structured	19,628
meta	the author's stance or purpose	17,923
meta	what kind of document it is	8,289
meta	what it leaves out	8,185
meta	who it is written for	5,183
meta	suggest a title	671

Sources

Category	docs1 sources	Sessions	% sessions	% tokens
code	stack-edu, stackexchange, rules, skillmd	607,071	30.7%	37.3%
general	finewiki, fineweb-edu, finepdfs-edu, news	514,942	26.0%	22.1%
science	pes2o, ar5iv, pubmed	284,680	14.4%	13.9%
docs	libretexts, pressbooks, oercommons, peps	163,348	8.3%	7.4%
legal	caselaw, usgpo, regulations, uk-hansard	160,540	8.1%	7.1%
transcript	youtube	136,914	6.9%	6.9%
literature	gutenberg, doab, public-domain-review	110,366	5.6%	5.4%

Turns, thinking and endings

90% of user turns have thinking on. Sessions have 1–8 user turns (agentic at most 3, reconstruction at most 2, dialogue at least 2).

Status	Sessions	Share
ok	1,954,550	98.82%
length	18,388	0.93%
tool_rounds	4,915	0.25%
context	8	0.00%

ok: every turn answered. length: a generation hit its token cap (8,192 per step), so the session ends at that message. tool_rounds: 8 tool rounds in one turn. context: the conversation outgrew 32K tokens.

Knobs

Tool use

Agentic sessions: 168,288. MiniCPM5 made 319,196 tool calls in them; 49.0% of tool results were not available (unknown tool or no mock left).

Fields

Field	Meaning
id	session id (uuid5)
doc_id, docs1_revision, doc_sha256	the docs1 document, pinned
source_type, bucket, doc_tokens	document source, length bucket, MiniCPM5 tokens
family, archetype	task family and archetype
knobs	JSON: voice, mention, asks_length, asks_format, system, difficulty, evidence spans
system, tools, mocks, turns	the skeleton (tools/mocks as JSON strings; turns with the thinking switch)
messages	JSON: the full session in the student view
status, finish	how the session ended; vLLM finish reason per generation
completion_tokens, requests	generated tokens and generation steps
generator, teacher	skeleton model, body model

Limitations

• No quality filter: the teacher is a 2B model sampled at temperature 1.0, so some answers are wrong, some thinking rambles, and some generations hit the length cap. Use status and the stats here to rebalance.
• Mock tools ignore arguments, so a result can fit the scenario better than the call that produced it.
• Luna skipped about 1.1% of planned slots and wrote a different turn count than planned for about 0.7%.
• Documents keep their docs1 licenses (see the docs1 card); generated text follows the same terms.

Code: datasets/qualia1-sft/ in the AlphaQualia repository (plan.py, skeleton.py, fill.py, card.py).