skycmd-route / README.md
maxgfr's picture
skycmd 2eb6c9b6: tool-routing slates (s6-route-nano, s6-route)
dab1725 verified
|
Raw History Blame Contribute Delete
20.3 kB
metadata
license: other
license_name: per-source
license_link: https://github.com/maxgfr/skycmd/blob/main/DATA_LICENSES.md
language:
  - en
  - fr
pretty_name: skycmd tool-routing slates (s6-route-nano, s6-route)
task_categories:
  - text-classification
tags:
  - skycmd
  - tool-routing
  - function-calling
  - ai-generated
configs:
  - config_name: tools_train
    data_files: data/tools_train.parquet
    default: true
  - config_name: intents_train
    data_files: data/intents_train.parquet
  - config_name: when2call_train
    data_files: data/when2call_train.parquet
  - config_name: tools_val
    data_files: data/tools_val.parquet
  - config_name: intents_val
    data_files: data/intents_val.parquet
  - config_name: when2call_val
    data_files: data/when2call_val.parquet
  - config_name: v_tools
    data_files: data/v_tools.parquet
  - config_name: v_tools_masked
    data_files: data/v_tools_masked.parquet
  - config_name: v_tools_pairs
    data_files: data/v_tools_pairs.parquet
  - config_name: v_tools_swapped
    data_files: data/v_tools_swapped.parquet
  - config_name: t_tools
    data_files: data/t_tools.parquet
  - config_name: t_tools_masked
    data_files: data/t_tools_masked.parquet
  - config_name: t_tools_pairs
    data_files: data/t_tools_pairs.parquet
  - config_name: t_tools_swapped
    data_files: data/t_tools_swapped.parquet
  - config_name: t_metatool
    data_files: data/t_metatool.parquet
  - config_name: t_clinc
    data_files: data/t_clinc.parquet
  - config_name: t_massive
    data_files: data/t_massive.parquet
  - config_name: t_when2call
    data_files: data/t_when2call.parquet
  - config_name: t_bfcl
    data_files: data/t_bfcl.parquet
  - config_name: f_tools
    data_files: data/f_tools.parquet
  - config_name: f_tools_masked
    data_files: data/f_tools_masked.parquet
  - config_name: f_metatool
    data_files: data/f_metatool.parquet
  - config_name: f_clinc
    data_files: data/f_clinc.parquet
  - config_name: f_massive
    data_files: data/f_massive.parquet
  - config_name: teacher_describe_intent
    data_files: data/teacher_describe_intent.parquet
  - config_name: teacher_rewrite_description
    data_files: data/teacher_rewrite_description.parquet

skycmd tool-routing slates (s6-route-nano, s6-route)

The training, validation and test records of the skycmd tool router: a user request, a list of up to 7 tools (each a name and one sentence) plus the option none of these tools fits the request, and the right option. Built from public tool-calling and intent sets under their own licences, and the descriptions the teacher wrote. The two models trained and tested on it:

  • s6-route-nano, a 2.4 MB ternary decision model: failed its gate (on T-tools 0.8395 against BM25's 0.8500, and 38.9 ms per request in WASM against a bar of 2 ms; docs/s6-nano2-prereg.md, result block);
  • s6-route, the same file fused with a 4-bit potion-base-2M embedding: passed its gate on fresh slates (0.8868 on T2-tools against BM25's 0.8169; docs/s6-nano3-prereg.md, result block).

Part of skycmd. Built by scripts/build-route-data.sh (the slates; docs/s6-nano2-prereg.md §4), scripts/route-teacher.sh (the teacher's jobs) and sky-data --seed 1235 route fresh (the fresh slates; docs/s6-nano3-prereg.md §4), assembled by scripts/hf-publish-data.py --only route at commit 2eb6c9b6. Licences of every set: DATA_LICENSES.md.

Machine-generated text. The teacher, GLM-5.3-Flash (zai-org/GLM-5.3-Flash, open weights, MIT), called through a hosted endpoint, wrote: the description of every intent option (MASSIVE, Banking77 and CLINC150 intents, written from the intent name alone: the configs intents_*, t_clinc, f_clinc, t_massive, f_massive and teacher_describe_intent), and the right tool's description in every renamed slate (teacher_rewrite_description). It also chose which hard distractors were replaced (the false-negative filter). It wrote no request, no label and no target, and it never saw a test request or a test slate. Upstream, MetaTool's queries were written by ChatGPT / GPT-4, and several tool-calling sets were generated with language models (see their cards).

Configs and splits

One config per file. split says what a file is for: train (the records the models were trained on), val (validation: the pilots, the calibration, the baselines' fit, the thresholds), test (s6-route-nano's tests, read once on 16/11/2026), fresh (s6-route's tests, built after that reading from requests no earlier file used, read once on 16/11/2026).

file split what it is rows bytes
data/tools_train.parquet train training records of the tool sources: four slates per request (natural, hard, removed, renamed) 81,596 13,343,233
data/intents_train.parquet train training records of MASSIVE and Banking77, the same four kinds 95,411 5,185,692
data/when2call_train.parquet train training records of When2Call (natural, hard, renamed) 14,621 2,600,986
data/tools_val.parquet val the four kinds on the validation side 4,185 515,195
data/intents_val.parquet val same 11,543 611,549
data/when2call_val.parquet val same 896 182,027
data/v_tools.parquet val validation slates built like T-tools (8 options); the pilots, the baselines' fit, the ask threshold 1,144 233,581
data/v_tools_masked.parquet val v_tools with every tool renamed tool_<k> 1,144 303,314
data/v_tools_pairs.parquet val 500 requests twice: gold present, gold removed 1,000 149,984
data/v_tools_swapped.parquet val the descriptions of the gold and of one distractor swapped 500 152,573
data/t_tools.parquet test T-tools: tools of held-out clusters, the gated test of s6-route-nano 2,000 445,205
data/t_tools_masked.parquet test T-tools with every tool renamed 2,000 608,875
data/t_tools_pairs.parquet test T-tools gold present / gold removed pairs 1,000 188,098
data/t_tools_swapped.parquet test T-tools with two descriptions swapped 500 183,556
data/t_metatool.parquet test T-metatool 2,000 160,952
data/t_clinc.parquet test T-clinc (in scope and out of scope) 2,000 93,251
data/t_massive.parquet test MASSIVE held-out scenarios, test partition 1,657 72,403
data/t_when2call.parquet test When2Call test/mcq reformatted 1,863 139,898
data/t_bfcl.parquet test BFCL multiple and irrelevance 440 53,421
data/f_tools.parquet fresh T2-tools: fresh requests about T-tools' held-out tools, the gated test of s6-route 945 219,764
data/f_tools_masked.parquet fresh T2-tools with every tool renamed 945 258,029
data/f_metatool.parquet fresh T2-metatool 2,000 158,885
data/f_clinc.parquet fresh T2-clinc 2,000 93,127
data/f_massive.parquet fresh T2-massive 2,000 85,177
data/teacher_describe_intent.parquet teacher one description per intent, written from its name alone 287 13,649
data/teacher_rewrite_description.parquet teacher one rewrite of each gold tool's description of the training and validation requests 14,263 1,067,284
files/pins.json — the repository's scripts/route/pins.json, copied — 4,579
files/MANIFEST.json — the data folder's MANIFEST.json, copied — 24,199
files/BUILD.json — the data folder's build/BUILD.json, copied — 28,164
files/FRESH.json — the data folder's build/fresh/FRESH.json, copied — 7,608
files/f_tools.clusters.jsonl — the data folder's build/fresh/f_tools.clusters.jsonl, copied — 77,415

Record schema

The slate configs share one schema (the typed-decision format of skycmd, docs/SPEC.md):

field type meaning
src string the group and source (tools:<source>, intents:<source>, when2call:when2call) in the training and validation records; the file (v-tools, t-clinc, f-massive-fr, ...) in the slate files
t string always choice
ins string the instruction, word for word: "Which tool should handle the user's request?"
opts list of strings the options: name: description for each tool (the description cut to its first sentence and at most 30 words), then none of these tools fits the request, always last
state string User: <request> (one user turn). In the training and validation records (*_train, *_val) a second line follows, BM25: <name> score <s> rank <r>; ..., the lexical score of each tool option against the request that the model reads; the slate files (v_*, t_*, f_*) carry no such line, and sky-data route render adds it, as the evaluation did
target list of floats one-hot: 1 on the right option
split string train, val, test or fresh
uid string the slate's identifier: the request's, with /<kind> in the training and validation records
kind string training and validation records: natural (the request's own tools, then random tools of its split), hard (the nearest tools by potion-base-2M, cleaned by the teacher), removed (the right tool taken out: the answer is none), renamed (every tool named tool_<k>); slate files: test (drawn by the test rule), removed, masked (every tool renamed), swapped (two descriptions swapped)
lang string en or fr
request string the request's identifier, <source>:<row>
tools list of strings the identifier of each tool option before none: t and 16 hex digits of the sha256 of its name: description as the source gives it
follow int swapped slates only: the index of the option that carries the right tool's description
source string the source of the request text (the prefix of request); its licence is the row's
option_sources list of strings every source whose data hold the text of one of the tool options (a distractor comes from another request of the same split, often of another source; one text can be in several sources)

teacher_describe_intent: id (<source>:<intent>), source, intent, description (the teacher's). teacher_rewrite_description: id (the tool identifier), name and original (the tool's name and description as the source gives them, the description cut and its key-like strings masked), rewrite (the teacher's, same meaning in other words; a renamed slate shows it cut to its first sentence and 30 words), sources (the sources that hold the tool).

Sources and licences

Each row keeps the licence of its sources: the request text is source's, the option texts are option_sources'; the teacher's descriptions are the outputs of an open-weight model under the MIT licence. Attribution goes to the upstream authors; the conversion is ours.

source upstream licence commit role
glaive glaiveai/glaive-function-calling-v2 apache-2.0 e7f4b645 train / val / T-tools / T2-tools: the first user turn answered by a call
hermes NousResearch/hermes-function-calling-v1 (func-calling-singleturn, func-calling, glaive-function-calling-5k) apache-2.0 dae3e1d2 same; a request whose normalised text is a glaive request's is dropped
toolace Team-ACE/ToolACE (data.json) apache-2.0 6bda777c same
seal-tools fairyshine/Seal-Tools on GitHub (Seal-Tools_Dataset/train.jsonl, tool.jsonl) apache-2.0 (LICENSE of the repository) ce753ecd same; its training file only
api-bank liminghao1630/API-Bank (training-data/lv1-api-train.json, lv2-api-train.json) mit (dataset card) 12e8158b same; ToolSearcher calls dropped
when2call nvidia/When2Call (train/*, through data/decisions/when2call.jsonl) cc-by-4.0 0582f774 train / val: a call to the called tool, no call to none
when2call-mcq nvidia/When2Call (test/when2call_test_mcq.jsonl) cc-by-4.0 0582f774 test only: the first 2,000 records reformatted
massive MASSIVE (FitzGerald et al., 2022) through mteb/amazon_massive_intent, en and fr (data/commands/massive.jsonl) cc-by-4.0 not pinned (sky-data fetch-commands) train / val: 14 of 18 scenarios; each intent a tool described by the teacher
massive-heldout MASSIVE, the 4 held-out scenarios (datetime, email, play, recommendation), en and fr cc-by-4.0 not pinned (sky-data fetch-commands) test: their test partition; fresh: their train and validation partitions
banking77 PolyAI/banking77 (data/decisions/banking77.jsonl) cc-by-4.0 not pinned (sky-data fetch-decisions) train / val only; each intent a tool described by the teacher
clinc clinc/clinc_oos (plus) cc-by-3.0 155b9c71 test only (T-clinc, T2-clinc); its 150 intent names described by the teacher
metatool HowieHwong/MetaTool on GitHub (all_clean_data.csv, plugin_des.json, Task2-Subtask1.json) mit (LICENSE); its queries were written by ChatGPT / GPT-4 upstream 35e81bb7 test only (T-metatool, T2-metatool)
bfcl-multiple gorilla-llm/Berkeley-Function-Calling-Leaderboard (BFCL_v3_multiple) apache-2.0 61fc0608 test only
bfcl-irrelevance gorilla-llm/Berkeley-Function-Calling-Leaderboard (BFCL_v3_irrelevance) apache-2.0 61fc0608 test only

Rows and requests per split, counted from the files above, and the requests the build kept for each side (build/counts/split.json; DATA_LICENSES.md gives these for training and validation and for the test side of the tool sources, and slates for the test-only sources):

source rows train rows val rows test rows fresh requests train / val / test / fresh requests kept by the build (train / val / test)
glaive 7,001 574 412 968 1,870 / 80 / 142 / 484 1,881 / 80 / 161
hermes 3,298 370 257 38 885 / 54 / 95 / 19 892 / 54 / 112
toolace 26,675 2,764 1,993 276 7,112 / 396 / 718 / 138 7,312 / 396 / 831
seal-tools 38,320 3,602 2,333 350 10,215 / 510 / 858 / 175 10,381 / 510 / 1,007
api-bank 6,302 663 505 258 1,677 / 104 / 187 / 129 1,910 / 104 / 209
when2call 14,621 896 — — 6,571 / 395 / — / — 12,801 / 699 / —
when2call-mcq — — 1,863 — — / — / 1,863 / — — / — / 2,000
massive 53,098 9,239 — — 14,162 / 2,462 / — / — 16,400 / 2,848 / —
massive-heldout — — 1,657 2,000 — / — / 1,657 / 2,000 — / — / 1,657
banking77 42,313 2,304 — — 11,281 / 616 / — / — 12,359 / 663 / —
clinc — — 2,000 2,000 — / — / 2,000 / 2,000 — / — / 5,500
metatool — — 2,000 2,000 — / — / 2,000 / 2,000 — / — / 20,614
bfcl-multiple — — 200 — — / — / 200 / — — / — / 200
bfcl-irrelevance — — 240 — — / — / 240 / — — / — / 240
all 191,628 20,412 13,460 7,890

Why the requests in the files are fewer than those kept by the build: a training or validation request is dropped when the teacher does not name its right tool as able in its hard slate (training / validation: tools 617 / 28, intents 3,316 / 433, When2Call 155 / 13; docs/s6-nano2-prereg.md amendment 2); When2Call's no-call slates are thinned at random to the 20% share of none (amendment 1; 14,570 training and 715 validation slates dropped), which leaves some requests with no slate; T-tools draws 2,000 slates from the 2,320 test-side requests, and the fresh T2-tools file uses the others and those over the cluster cap; T-clinc and T-metatool draw 2,000 slates each from their sources; 137 of the 2,000 When2Call test items list no tool, so no slate (docs/s6-nano2-readings.md, item 34). A request is in up to four training slates and in several reading-check files.

Not in this repository:

  • build/test/t_toolrouting.jsonl: the 100 test slates of stanley-nv/tool-routing-data, whose card asks users to review the generating model's terms before training other models; scripts/build-route-data.sh rebuilds them.
  • build/val/v_tools.model.jsonl: the model's copy of v_tools (the BM25 line added to each state); sky-data route render rebuilds it.
  • build/val/v_tools_masked.model.jsonl: same, for v_tools_masked.
  • build/val/v_tools_pairs.model.jsonl: same, for v_tools_pairs.
  • build/val/v_tools_swapped.model.jsonl: same, for v_tools_swapped.
  • build/teacher/describe-intent.items.jsonl: the items sent to the teacher (the intent names, also in the id of each description).
  • build/teacher/false-negatives.items.jsonl: the items of the false-negative filter (training and validation hard slates before filtering); sky-data route split rebuilds them.
  • build/teacher/false-negatives.jsonl: the teacher's answers of the false-negative filter (option indices into the items above); their effect is in the hard slates.
  • build/teacher/rewrite-description.items.jsonl: the items sent to the teacher (the tools' own descriptions, given in original).
  • build/teacher/describe-intent.refused.jsonl: items refused by the endpoint's moderation (empty).
  • build/teacher/false-negatives.refused.jsonl: items refused by the endpoint's moderation.
  • build/teacher/rewrite-description.refused.jsonl: items refused by the endpoint's moderation (empty).
  • the raw upstream files (src/, hf-cache/), the models (models/, embed/), the Python environment (envs/), the embeddings, stage files and evaluation outputs of the data folder: scripts/build-route-data.sh downloads every source at the commit of files/pins.json, checks each file against the hash its host publishes (files/MANIFEST.json: URL, commit, bytes, sha256, licence) and rebuilds every file here byte for byte (files/BUILD.json, files/FRESH.json: the hashes and counts of the build).

Limits

  • English mostly: 32,365 of the 233,390 slate rows are French, all from MASSIVE (requests in French, intent names and descriptions in English).
  • Synthetic upstream data: several tool-calling sets and MetaTool's queries were written by language models upstream; the build filters and deduplicates, it does not check them by hand. The share of validation slates that hold a second valid tool (the label-noise estimate of docs/s6-nano2-prereg.md §4.5) is not in the result files.
  • Test sets partly in distribution: T-tools holds tools of clusters never trained on ("fresh, tool-disjoint: no test tool cluster in training"), but from the same five sources as the training tools; "the clustering that defines 'unseen tool' uses all-MiniLM-L6-v2, which is also a reported baseline". "The When2Call test and MASSIVE test utterances were read in v1's evaluation" (docs/s6-nano2-prereg.md §10); When2Call and the other MASSIVE scenarios are in training. On T2-tools "these requests are about the same held-out tools as v2's test", and 107 of its 945 slates repeat the normalised request of an earlier slate (docs/s6-nano3-prereg.md §7 and amendment 1). MetaTool, CLINC150 and BFCL are never trained.
  • Keys: the builder masks key-like strings in requests and descriptions as <KEY> (sky_data::route::mask_keys; 280 in requests, and 103 tools held one in their schema or description). Every string of data/ and files/ was scanned with the same rule at assembly, and none is left. The builder does not mask tool names: five tool names, and six repository paths in files/, match the rule without being keys; they are listed by hash in scripts/hf-publish-data.py (ROUTE_NOT_KEYS).
from datasets import load_dataset
ds = load_dataset("maxgfr/skycmd-route", "tools_train")