semantic-repair-routing

The 84,819 supervised pairs that trained SemanticRepair-270M: a message somebody actually wrote, and the requests inside it restated plainly, one per line.

It teaches one narrow thing. An embedding router compares a question with the description of every capability it can reach. People do not write the way capabilities are described — they hedge, they apologise, they ask two things in one breath, they name what they do not want. This data pairs the first with the second, so a small model can restate a question into the form the router was built to compare against.

The files

file	rows
train.jsonl	82,819
valid.jsonl	1,000
test.jsonl	1,000

One JSON object per line, two keys, no nesting:

{"prompt": "Spanish: no busques la tienda de ropa, dime la dosis de paracetamol\n=>\n",
 "completion": "dime la dosis de paracetamol"}

prompt is the English name of the language, a colon, the message verbatim, then a newline, =>, a newline. completion is one request per line, or the single token NO_REQUEST.

The completion surface is bare on purpose: no chat template, no system message, no examples. A model trained here and then served through a chat template is being asked a different question than it was taught.

What is in it

Five languages, close to balanced:

	rows
English	18,006
Spanish	16,903
German	16,659
French	16,602
Italian	14,649

12.6% teach the model to refuse. 10,432 rows answer NO_REQUEST: somebody said something that is not a request at all. A rewriter that always rewrites turns small talk into a routed action, which is the failure that costs the most.

16.9% carry more than one request. 14,028 rows answer with two or more lines, because one message asking for two things is the case a router handles worst and the case people write most naturally.

What each message is doing. Every one of the 82,819 training rows carries one of 38 labels naming the kind of case it is. This is the whole table, because the shape of the data is the point of the data:

family	rows
massive_real	5,736	6.9%
minimal_pair	5,299	6.4%
indirect	4,912	5.9%
presto_native	4,606	5.6%
oasst_human	4,450	5.4%
multi	3,974	4.8%
direct	3,904	4.7%
vent_statement_request	3,516	4.2%
negation_exclusion	3,398	4.1%
trap_topic_lure	3,324	4.0%
statement	3,230	3.9%
conditional	2,815	3.4%
long_rambling	2,776	3.4%
typo_slang	2,629	3.2%
noun_phrase	2,547	3.1%
injection_command	2,410	2.9%
vent	2,375	2.9%
polite_indirect	2,325	2.8%
self_correction	2,323	2.8%
multi3	2,305	2.8%
injection_meta	2,278	2.8%
code_switching	1,934	2.3%
ultra_short	1,930	2.3%
multi2	1,483	1.8%
graph3	1,169	1.4%
multi4	864	1.0%
graph2	721	0.9%
graph4	552	0.7%
a weekday in the past, like 'last Tuesday'	500	0.6%
a named holiday, like Easter or Christmas	447	0.5%
a period like 'last week', 'last month', 'this year'	444	0.5%
yesterday, today or tomorrow	314	0.4%
an explicit day and month, like '18 July'	307	0.4%
a range between two days or two months	307	0.4%
a month by name	298	0.4%
a phrase that bounds and orders at once, like 'the last meeting before Easter'	296	0.4%
wave1_multi	73	0.1%
meta	48	0.1%

massive_real and oasst_human are real traffic; presto_native is natively written rather than translated. minimal_pair holds two messages differing in one word, so a model cannot pass by matching topics. trap_topic_lure names a topic prominently and asks for another, to pull routing the wrong way. The eight date families are the hardest column and the model trained here is still weakest on them.

The two families that decide what this model refuses to do

injection_meta — 2,278 rows — is a message aimed at the assistant itself: reveal your instructions, ignore what you were told, pretend you have no rules. 1,901 of them answer NO_REQUEST, because there is nothing here to route. Together with injection_command these two families supply 19% of every NO_REQUEST row in the set: refusing is largely taught by messages that try hardest to get an answer.

injection_command — 2,410 rows — is the opposite lesson, and the more delicate one. A hostile but explicit command about the user's own things: delete every saved bank account, empty my personal cloud, cancel all my bookings tonight. 2,281 of them are restated faithfully, action intact.

The rewriter does not sanitise and does not soften. Deciding whether an action is allowed is not its job and it has no way to do that job well: it sees one sentence, not the account, the permissions or the consequences. A model that quietly dropped "delete" would hand the router a different request from the one the person made, and the person would never learn their instruction had been edited. So the restatement carries the command as written, the router places it, and whatever runs it is where refusing belongs.

How it was built, and what was thrown away

Harvested from 53 files of generated and judged messages, then filtered hard. The funnel, reproduced by rerunning the real code rather than reading it:

	rows
read	540,344
duplicate messages	301,079
no message, or no requests attached	82,750
judged REJECT	72,996
kept	83,519

Then one more filter, and it is the one worth knowing about: 1,631 rows were removed for teaching the model to lose a date. In a message naming a day, a completion that drops the day teaches exactly the failure the seat is measured on. 1,586 of those genuinely lost the day; the rest were abstentions that only looked like it.

Only 20 of the 53 files contribute a single surviving row. The other 33 were read for 249,105 rows and gave nothing: thirteen carry no requests field at all, the rest were duplicates or rejected.

Which models wrote it, and which judged it

Every row records the model that generated it. All 82,819 are accounted for.

generator	rows
Qwen3.6-35B	72,697	87.8%
Qwen3.8-27B	7,935	9.6%
Gemma-4-E4B	2,187	2.6%

the two roles
generator	samples at its own published preset
judge	greedy, temperature 0, seeded: a verdict does not change between runs
judge answers	a two-word verdict against written rules, not an opinion
rejected outright	72,996 rows, not in this set
after the judge	deduplication, field checks, the date filter: ordinary code, no model votes

For most of this data the judge was a different model from the generator, which is what makes a verdict worth something. That did not hold to the end: two of the three generators and the Gemma-4-31B judge were retired from the machine in late August 2026, leaving one model doing both jobs. What still separates the roles is the decoding and the fixed verdict. It is a weaker guarantee than two models, and it is stated here rather than left to be discovered.

Where all of this happened

One desktop machine, and nothing left it.

machine	Mac mini, Mac16,11
chip	Apple M4 Pro
CPU	14 cores, 10 performance and 4 efficiency
GPU	20 cores
memory	64 GB unified
system	macOS 26.4

the fine-tune
method	mlx_lm lora, full fine-tune
iterations	20,704
tokens seen	2,383,728
peak memory	7.3 GB
speed	5.6 iterations/s, 610 tokens/s
wall clock	11:15 to 12:42, about 90 minutes

Generators and judge ran on that same machine as local llama.cpp servers on loopback. No message here was written by a hosted API, and none was sent to one to be judged. The unified memory is why: a 35B generator and a fine-tune only share a desk when CPU and GPU read the same 64 GB.

A model that decides where a question goes does not need a cluster or a month. It needs a narrow job, data built for it, and a bench honest enough to say when a round did not help.

What it is not

Not answers. No completion here answers anything. The restatement is compared against capability descriptions; whatever runs, runs on the original message.

Not a benchmark. valid and test are held out from the same distribution and were used to gate training. Judging a model's routing on them measures agreement with this data, not routing quality — for that the model repository reports two independent benches over different corpora.

Not human-written throughout. The messages were generated and then judged, with rejected rows dropped rather than kept and corrected. What survives is what a judge accepted, which is a real limit: a bias in the generator is a bias here.

Known limits

Dates are the weak column. Even after removing the rows that taught the loss, the model trained on this data drops the named day in about a third of dated questions. The data does not carry enough dated messages to fix that by itself.

Five languages, and the two the engine has contracts for are the ones measured. French, German and Spanish are present in the data and judged against English bars, which is not the same as being calibrated.

Licence

CC BY 4.0. The model trained on it is under the Gemma terms, which travel with the weights and not with this data.