ihemanthc/tollgate-routing-data

Tollgate routing labels

Each row is one user query labelled with its minimum sufficient tier: the cheapest of three model tiers (local_small, mid_tier, frontier) whose answer an LLM judge rated as good as the frontier tier's answer. Tollgate fine-tunes a router on these labels to send each query to the cheapest model that can handle it.

Blocks marked tollgate:begin/tollgate:end are generated by tollgate push-dataset from the code and the data. No number in this card is written by hand.

What is and isn't included

Source	Rows	Rows with text	Rows as sha256 + labels only
gsm8k	105	105	0
lmsys-chat-1m	91	0	91
mmlu	104	104	0
all	300	209	91

Included

• dataset.parquet: one row per labelled query, with the columns below.
• splits/train.txt, splits/calibration.txt, splits/test.txt: the query_ids of each split, one per line, sorted. They match the split column.
• rebuild_dataset.py: restores withheld query text for users who have accepted the source's license (see Rebuilding withheld text).
• Query text (prompt) only for sources whose license permits redistribution.

Not included

• Any query text, conversation text or summary from a source that forbids redistribution (LMSYS-Chat-1M). Those rows carry only text_sha256 and Tollgate's labels.
• Any other field of a source dataset: LMSYS conversation ids, model names, later turns, language tags and OpenAI moderation outputs; GSM8K answers and solutions; MMLU answer keys.
• The answers produced by the three tiers, the judge's raw output and its one-line reasons, the per-call cost ledger and the response cache.
• Any user or account metadata.

Columns:

Column	Contents
query_id	sha256 of the exact query text (UTF-8); the row key.
text_sha256	Same value as query_id, named for what it is: how withheld text is matched.
text_included	True when prompt carries the text; False when the source forbids it.
prompt	Query text; empty (null) for rows from a source that forbids redistribution.
conversation_summary	Prior-turn summary; empty for withheld sources and single-turn rows.
source	Seed source tag (see Seed sources).
split	train / calibration / test.
tier	Label: minimum sufficient tier (local_small, mid_tier or frontier).
needs_tools	Heuristic flag from the frontier answer's text; placeholder, not judged.
needs_rag	Heuristic flag from the frontier answer's text; placeholder, not judged.
judged_tiers	Cheaper tiers that received a judge verdict for this query.
frontier_cost_usd	What the frontier tier's answer cost.
label_cost_usd	What the answer at the label tier cost.

Seed sources and licenses

Queries are drawn round-robin from three sources, at the pinned revisions below, after dropping empty prompts, prompts over 1500 characters, and duplicates (whitespace- and case-insensitive).

Source	Hub dataset @ revision	License	Query text published?	Basis
lmsys-chat-1m	lmsys/lmsys-chat-1m @ 200748d9d3cd	LMSYS-Chat-1M Dataset License Agreement	no: sha256 + labels only	"Prohibited Transfers: You should not distribute, copy, disclose, assign, sublicense, embed, host, or otherwise transfer the dataset to any third party."
gsm8k	openai/gsm8k @ 740312add88f	MIT	yes	MIT permits redistribution with the copyright and permission notice (below).
mmlu	cais/mmlu @ c30699e8356d	MIT	yes	MIT permits redistribution with the copyright and permission notice (below).

Any source not listed here is treated as non-redistributable: its text is withheld.

MIT notice for the GSM8K and MMLU text included in prompt:

Copyright (c) 2021 OpenAI
Copyright (c) 2020 Dan Hendrycks

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and
associated documentation files (the "Software"), to deal in the Software without restriction,
including without limitation the rights to use, copy, modify, merge, publish, distribute,
sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or
substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT
NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT
OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Cite the sources you use:

• Zheng et al., LMSYS-Chat-1M (2023), arXiv:2309.11998
• Cobbe et al., Training Verifiers to Solve Math Word Problems (2021), arXiv:2110.14168
• Hendrycks et al., Measuring Massive Multitask Language Understanding (2021), arXiv:2009.03300

LMSYS-Chat-1M's agreement also lets its authors require deletion of the dataset "and related materials", and it forbids attempting to identify individuals in it. If they ask for deletion, the LMSYS-derived rows here (hashes and labels) will be removed.

Labelling procedure

1. Answers. Every query is answered once by each tier at temperature 0, capped at 4096 output tokens (reasoning tokens included); the judge has the same cap. Calls are cached by sha256(prompt, model), so no query is answered twice by the same model, except that a reply which ran into an earlier, lower cap is asked again under the higher one.

2. Judging. For each cheaper tier (local_small, mid_tier), an LLM judge sees the question, the frontier answer as the reference and the cheaper answer as the candidate. It returns one of:

   • equivalent: same substance as the reference (same final answer, key facts or steps); only wording, length or formatting differ.
   • acceptable: different or less thorough, but correct and complete enough that the user would be satisfied.
   • worse: wrong, missing something the user needs, unsafe, off-topic, or a refusal the reference did not make.

   Output is forced to JSON and validated. One retry is allowed on unparseable output. A verdict that still fails is recorded as missing, never guessed.

3. Label. tier is the cheapest tier judged equivalent or acceptable, or frontier if none was. A cheap tier that passes wins even if a pricier tier was judged worse.

4. Dropped rows. A query is left out if its frontier answer failed, or if any tier cheaper than its label has no verdict. Such a tier might have been sufficient, so its label could be too expensive.

5. needs_tools / needs_rag. These are not judged. They are regex heuristics over the frontier answer (tool-call markup, "I can't browse", knowledge-cutoff phrasing) and are placeholders until hand labels replace them.

Models

Role	Model(s)
local_small	ollama/qwen2.5:7b
mid_tier	ollama/qwen2.5:14b
frontier	ollama/qwen2.5:32b
judge	ollama/mistral-small:24b

The judge is never the frontier model

The judge must not be the model that wrote the frontier answer. The reference in every comparison is the frontier answer, so a frontier-model judge would be grading candidates against its own output. LLM judges prefer their own generations (self-preference). That would mark more cheap answers worse, push labels up to frontier, and train a router that overspends.

Models are compared by base name (openai/gpt-4o and azure/gpt-4o are the same model). The rule is enforced when collecting (tollgate collect-all, tollgate judge) and again when publishing: tollgate push-dataset refuses if any verdict for these rows came from a frontier model. It does not exclude other models from the frontier model's family or vendor; see below.

Known judge biases

None of these has been measured on this dataset yet. They are the known failure modes of single-judge LLM labelling, applied to this setup.

• Position. The reference is always shown before the candidate, never in swapped order, so any position preference of the judge is uncorrected.
• Family and self-preference. The judge may favour answers from its own model family or vendor. If it shares a family with the mid tier, mid_tier labels may be inflated.
• Verbosity. LLM judges tend to rate longer answers higher. Cheaper tiers often answer more briefly, which pushes labels up.
• Reference anchoring. Candidates are judged against the frontier answer. The prompt asks for judgement against the question, but a correct cheap answer that disagrees with a wrong reference can still be rated worse.
• The acceptable boundary. It is subjective, and it counts as sufficient: a lenient judge pulls labels down.
• One judgement per pair. There is a single sample at temperature 0, no second judge, and no human audit yet.
• Missing verdicts. Queries whose verdicts failed to parse can be dropped (step 4 above), which can bias the sample toward questions that are easy to judge.
• Language. LMSYS prompts span many languages, and judge quality outside English is less certain.

Label distribution

Label (minimum sufficient tier)	Rows	Share
local_small	233	77.7%
mid_tier	40	13.3%
frontier	27	9.0%

Heuristic flag	Rows true	Share
needs_tools	0	0.0%
needs_rag	0	0.0%

Label by source:

Source	local_small	mid_tier	frontier	total
gsm8k	95	8	2	105
lmsys-chat-1m	63	13	15	91
mmlu	75	19	10	104

Splits

Rows are split 70/15/15 into train, calibration and test, stratified by (label, source) with a fixed seed. Each stratum splits to within one row. The calibration split is reserved for fitting the router's temperatures and is never used in training. Early stopping uses a holdout carved out of train.

Split	local_small	mid_tier	frontier	total	Share of rows
train	163	29	19	211	70.3%
calibration	34	6	4	44	14.7%
test	36	5	4	45	15.0%

Rebuilding withheld text

For rows with text_included = false, text_sha256 is sha256 of the exact UTF-8 query text, as it was extracted: for LMSYS-Chat-1M, the first user message of a conversation, with leading and trailing whitespace stripped. With access to the source granted under its own license:

pip install pandas pyarrow datasets
huggingface-cli login
python rebuild_dataset.py --dataset dataset.parquet --out dataset.rebuilt.parquet --limit 0

The script streams the source at the revision listed above and fills prompt for every hash it finds. --limit 0 scans the whole source. The default --limit is a quick smoke test.

Terms that still apply to you

The labels come from outputs of the tier and judge models listed under Models. Their providers' terms of use may limit how you use model-derived data. LMSYS-derived rows remain subject to the LMSYS-Chat-1M agreement for anyone who rebuilds their text.

Generated 2026-10-06 14:29 UTC by tollgate 0.1.0 from 300 labelled rows.