skycmd-route7 / README.md
maxgfr's picture
skycmd c8799b5a: the data of s6-route v7 (s6-route7-nano, gate failed)
ded4ea0 verified
|
Raw History Blame Contribute Delete
26.5 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 data of s6-route v7 (s6-route7-nano, gate failed)
task_categories:
  - text-classification
tags:
  - skycmd
  - tool-routing
  - function-calling
  - ai-generated
  - negative-result
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: v5_tools
    data_files: data/v5_tools.parquet
  - config_name: v5_tools_model
    data_files: data/v5_tools_model.parquet
  - config_name: v5_tools_masked
    data_files: data/v5_tools_masked.parquet
  - config_name: v5_tools_masked_model
    data_files: data/v5_tools_masked_model.parquet
  - config_name: v5_tools_rw
    data_files: data/v5_tools_rw.parquet
  - config_name: v5_tools_rw_model
    data_files: data/v5_tools_rw_model.parquet
  - config_name: v5_tools_masked_rw
    data_files: data/v5_tools_masked_rw.parquet
  - config_name: v5_tools_masked_rw_model
    data_files: data/v5_tools_masked_rw_model.parquet
  - config_name: v5_tools_blind
    data_files: data/v5_tools_blind.parquet
  - config_name: v5_tools_blind_model
    data_files: data/v5_tools_blind_model.parquet
  - config_name: v5_tools_masked_blind
    data_files: data/v5_tools_masked_blind.parquet
  - config_name: v5_tools_masked_blind_model
    data_files: data/v5_tools_masked_blind_model.parquet
  - config_name: v4_tools_model
    data_files: data/v4_tools_model.parquet
  - config_name: v4_tools_masked_model
    data_files: data/v4_tools_masked_model.parquet
  - config_name: v_tools_model
    data_files: data/v_tools_model.parquet
  - config_name: v_tools_masked_model
    data_files: data/v_tools_masked_model.parquet
  - config_name: r_tools_natural_model
    data_files: data/r_tools_natural_model.parquet
  - config_name: r_tools_renamed_model
    data_files: data/r_tools_renamed_model.parquet
  - config_name: t5_tools
    data_files: data/t5_tools.parquet
  - config_name: t5_tools_masked
    data_files: data/t5_tools_masked.parquet
  - config_name: t5_tools_rw
    data_files: data/t5_tools_rw.parquet
  - config_name: t5_tools_masked_rw
    data_files: data/t5_tools_masked_rw.parquet
  - config_name: t5_tools_blind
    data_files: data/t5_tools_blind.parquet
  - config_name: t5_tools_blind_model
    data_files: data/t5_tools_blind_model.parquet
  - config_name: t5_tools_masked_blind
    data_files: data/t5_tools_masked_blind.parquet
  - config_name: t5_tools_masked_blind_model
    data_files: data/t5_tools_masked_blind_model.parquet
  - config_name: t5_tools_servers
    data_files: data/t5_tools_servers.parquet
  - config_name: teacher_rewrite_description_v4
    data_files: data/teacher_rewrite_description_v4.parquet
  - config_name: teacher_rewrite_description_t5
    data_files: data/teacher_rewrite_description_t5.parquet

skycmd tool-routing data of s6-route v7 (s6-route7-nano, gate failed)

The training, validation and test data of s6-route v7, the seventh and last version 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. Every slate file here is a file that v7's training, pilot, fit or gate read, converted to parquet row for row, each row with its sources; the teacher's rewrites that those slates show are here too.

The model trained on it is s6-route7-nano, and its pre-registered gate FAILED (docs/s6-route7-prereg.md, result block of 19/11/2026: G1 fails, G2 fails, G3 fails, G4 holds, G5 holds, G6 fails). On T5-tools the pair (the model part fused with a 4-bit potion-base-2M embedding) reads 0.5610 against BM25's 0.5675; with the request hidden, its model part is at chance (0.1406, chance 0.1429) and the pair is not (0.1969, bar 0.1929). The route line is closed with this result; the published router stays s6-route, whose data is skycmd-route.

The fresh test T5 is in this repository. It was opened once, on 19/11/2026 (files/FRESH-OPEN, which holds the hashes of the model, the embedding file and the fit read on it), read by v7's gate, and nothing will be gated on it again: the route line has no v8. Publishing it harms no future reading. Anyone who reports a new system on it should say that it is no longer fresh.

Part of skycmd. Built by scripts/build-route4-data.sh (v4: the teacher's rewrites and the Toucan fetch), scripts/build-route5-data.sh (v5: the surface cell, the records and T5's rule), scripts/build-route6-data.sh (v6: the When2Call fix, the blind copies, T5) and scripts/build-route7-data.sh (v7: the anchor of the no-call slates, docs/s6-route7-prereg.md §4.8), with scripts/route-teacher.sh for the teacher's jobs; trained by scripts/train-route7.sh and gated by scripts/route7-gate.sh. The pre-registrations: docs/s6-route4-prereg.md, docs/s6-route5-prereg.md, docs/s6-route6-prereg.md, docs/s6-route7-prereg.md. Assembled by scripts/hf-publish-data.py --only route7 at commit c8799b5a. 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 (the intents group, written from the intent name alone; its file is skycmd-route's teacher_describe_intent); the rewrite that replaces a tool's description with probability 0.5 in every kind of training and validation slate (D1; the options flagged in rewritten; v2's rewrites, skycmd-route's teacher_rewrite_description, and v4's, teacher_rewrite_description_v4); the rewritten distractor of the probe copies (*_rw; teacher_rewrite_description_t5 for T5's). It also chose which hard distractors were replaced (the false-negative filter of v5, v6 and v7). It wrote no request, no label and no target, and it never saw a test request or a test slate: for T5 it saw the tools' names and descriptions only. Upstream, requests were written by language models too: Toucan's by Mistral-Small, DevStral-Small, GPT-OSS, Kimi-K2 and Qwen3-32B (its card), 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 model was trained on), val (the trainer's validation, the pilot's rule, the calibration, the fit of beta and the ask threshold, the reported readings), fresh (T5, the gated test, read once). In the records, a row is one slate; a request is in up to four (natural, hard, removed, renamed). The *_model configs are the copies the model part read: each tool option starts with its BM25 score and rank (score <s> rank <r>; <name>: <description>, the render scores-in-options); the slate configs without _model are what the pair and the baselines read. ? is the placeholder of the blind copies, in place of the request (B0: the model part reads every BM25 score at 0.00; B1: the bare slates).

file split what it is rows bytes
data/tools_train.parquet train training records of the tool sources: natural, hard, removed, renamed (v5's surface cell, D1's rewrites); read by sky-train s2-decide, weight 0.60 77,070 13,901,637
data/intents_train.parquet train training records of MASSIVE and Banking77, the same four kinds; weight 0.25 91,130 6,611,154
data/when2call_train.parquet train training records of When2Call (natural, hard, renamed), its no-call slates rebuilt by v6 and v7; weight 0.15 13,885 2,704,821
data/tools_val.parquet val the four kinds on the validation side; the trainer's validation 3,941 574,022
data/intents_val.parquet val same 11,085 787,945
data/when2call_val.parquet val same 847 197,206
data/v5_tools.parquet val 1,144 validation slates by v5's test rule (915 tool-right, 229 none-right): the pair's fit of beta and tau_ask7, the export's tau_ask, the pair's validation readings 1,144 223,779
data/v5_tools_model.parquet val the model's copy of v5_tools (each tool's BM25 score and rank in its option): the calibration and every checkpoint's reading 1,144 277,916
data/v5_tools_masked.parquet val v5_tools with every tool renamed tool_<k> 1,144 271,937
data/v5_tools_masked_model.parquet val its model copy 1,144 299,537
data/v5_tools_rw.parquet val v5_tools with one random distractor's description replaced by its rewrite (follow): the pilot's rule 2 1,144 236,723
data/v5_tools_rw_model.parquet val its model copy 1,144 297,531
data/v5_tools_masked_rw.parquet val the masked copy with one distractor rewritten 1,144 284,242
data/v5_tools_masked_rw_model.parquet val its model copy 1,144 311,724
data/v5_tools_blind.parquet val B0 and B1 of v5_tools: the request replaced by ? (the pair in B0; the model part in B1 reads the same bytes, v5_tools.blind-bare.model.jsonl) 1,144 93,091
data/v5_tools_blind_model.parquet val B0 of v5_tools for the model part: every BM25 score 0.00 (the pilot's rules 3 and 4) 1,144 123,629
data/v5_tools_masked_blind.parquet val B0 and B1 of v5_tools_masked 1,144 141,249
data/v5_tools_masked_blind_model.parquet val B0 of v5_tools_masked for the model part 1,144 156,775
data/v4_tools_model.parquet val the model copy of v4's validation slates (the same 1,144 requests by v2's test rule): a reported reading at every checkpoint 1,144 277,568
data/v4_tools_masked_model.parquet val its masked copy 1,144 298,025
data/v_tools_model.parquet val the model copy of v2's v_tools (skycmd-route): a reported reading at every checkpoint 1,144 314,138
data/v_tools_masked_model.parquet val its masked copy 1,144 331,743
data/r_tools_natural_model.parquet val the natural validation records of the tools group, one per request: a reported reading 1,126 263,301
data/r_tools_renamed_model.parquet val the renamed ones: a reported reading 1,126 327,795
data/t5_tools.parquet fresh T5-tools: 2,000 fresh slates about the tools of 262 MCP servers (1,600 tool-right, 400 none-right), the gated test of s6-route7-nano 2,000 460,390
data/t5_tools_masked.parquet fresh T5-tools with every tool renamed (G2 (b), (c)) 2,000 583,208
data/t5_tools_rw.parquet fresh probe copy: the 1,600 tool-right slates with one distractor's description rewritten by the teacher (follow; G2 (a)) 1,600 405,875
data/t5_tools_masked_rw.parquet fresh probe copy of the masked file 1,600 505,334
data/t5_tools_blind.parquet fresh B0 and B1 of T5-tools, the request replaced by ? (G6; B1 reads the same bytes, t5_tools.blind-bare.model.jsonl) 2,000 200,468
data/t5_tools_blind_model.parquet fresh B0 of T5-tools for the model part (G6 (a)) 2,000 269,176
data/t5_tools_masked_blind.parquet fresh B0 and B1 of the masked file 2,000 323,286
data/t5_tools_masked_blind_model.parquet fresh B0 of the masked file for the model part 2,000 370,075
data/t5_tools_servers.parquet fresh the MCP server and the tool cluster of each T5-tools slate, by uid (the servers bootstrap of G1 and G6) 2,000 53,395
data/teacher_rewrite_description_v4.parquet teacher v4's 13,687 rewrites: one per tool of the training and validation pools and of the intents that had none (docs/s6-route4-prereg.md §4.2), shown in place of a description with probability 0.5 (D1) 13,687 1,002,672
data/teacher_rewrite_description_t5.parquet teacher one rewrite of each of the 1,448 Toucan tools that a T5 slate shows, from its name and description only: the probe copies t5_tools_rw and t5_tools_masked_rw, nothing else 1,448 123,744
files/inputs6.json — the repository's runs/s6-route7/inputs6.json, copied — 6,580
files/FRESH-OPEN — the repository's runs/s6-route7/FRESH-OPEN, copied — 1,086
files/t5-counts.json — the repository's runs/s6-route7/t5-counts.json, copied — 1,687
files/MANIFEST4.json — the data folder's MANIFEST4.json, copied — 2,251
files/v5-plan.json — the data folder's build7/counts/v5-plan.json, copied — 18,348
files/v5-slates.json — the data folder's build7/counts/v5-slates.json, copied — 18,035

Read by v7 and not repeated here:

  • build7/val/v5_tools.blind-bare.model.jsonl: the same bytes as build7/val/v5_tools.blind.jsonl (checked by sha256).
  • build7/val/v5_tools_masked.blind-bare.model.jsonl: the same bytes as build7/val/v5_tools_masked.blind.jsonl (checked by sha256).
  • toucan5/fresh/t5_tools.blind-bare.model.jsonl: the same bytes as toucan5/fresh/t5_tools.blind.jsonl (checked by sha256).
  • toucan5/fresh/t5_tools_masked.blind-bare.model.jsonl: the same bytes as toucan5/fresh/t5_tools_masked.blind.jsonl (checked by sha256).
  • build7/fix/val/v5_tools.jsonl: the same bytes as build7/val/v5_tools.jsonl (checked by sha256).
  • build7/fix/val/v5_tools_masked.jsonl: the same bytes as build7/val/v5_tools_masked.jsonl (checked by sha256).
  • build7/fix7/val/v5_tools.jsonl: the same bytes as build7/val/v5_tools.jsonl (checked by sha256).
  • build7/fix7/val/v5_tools_masked.jsonl: the same bytes as build7/val/v5_tools_masked.jsonl (checked by sha256).
  • build/teacher/describe-intent.jsonl (the intent descriptions of the intents group): skycmd-route's config teacher_describe_intent, the same file (its sha256 checked against the data folder's build/BUILD.json).
  • build/teacher/rewrite-description.jsonl (v2's rewrites of the gold tools, reused by D1): skycmd-route's config teacher_rewrite_description, the same file (its sha256 checked against the data folder's build/BUILD.json).
  • build/fresh/f_tools.jsonl (T2-tools, read for G5): skycmd-route's config f_tools, the same file (its sha256 checked against the data folder's build/fresh/FRESH.json).

Record schema

The slate configs share skycmd-route's schema (src, t, ins, opts, state, target, split, uid, kind, lang, request, tools, follow, source, option_sources; see its card) and add the fields of v5's builder:

field type meaning
rewritten list of bools records only: for each tool option, whether its description is the teacher's rewrite (D1)
cell_level int the level of v5's surface cell (L0 to L6) that the slate's options were drawn in (docs/s6-route5-prereg.md §4.3)
roles list of strings for each tool option, how it came into the slate: right (the right tool), exempt (one of the request's own tools, kept outside the cell rule, v5 §4.3) or drawn (drawn within the cell)
from string removed slates only: the kind of slate the right tool was taken out of (natural or hard)

In T5, lang is und: Toucan gives no language. t5_tools_servers: uid, server (the MCP server that Toucan gives for the request), cluster (the tool cluster of v2's rule), source. The teacher configs: id (the tool identifier), name and original (the tool's name and description as the source gives them), rewrite (the teacher's; a slate shows it cut to its first sentence and 30 words), sources (the sources that hold the tool), and for T5 servers.

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 texts are the outputs of an open-weight model under the MIT licence. Attribution goes to the upstream authors; the conversion is ours. Every source has its row in DATA_LICENSES.md, and the script refuses a row whose sources have none.

source upstream licence commit role in v7
glaive glaiveai/glaive-function-calling-v2 apache-2.0 e7f4b645 train / val (the tools group): 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 (the When2Call group): a call to the called tool, no call to none
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 (the intents group): 14 of 18 scenarios, en and fr; each intent a tool described by the teacher
banking77 PolyAI/banking77 (data/decisions/banking77.jsonl) cc-by-4.0 not pinned (sky-data fetch-decisions) train / val (the intents group); each intent a tool described by the teacher
bfcl-multiple gorilla-llm/Berkeley-Function-Calling-Leaderboard (BFCL_v3_multiple) apache-2.0 61fc0608 no request: the same tool texts are in some When2Call options (60 training and 6 validation rows of when2call_*)
toucan Agent-Ark/Toucan-1.5M (config SFT, SFT/train-0000{0,1,2}-of-00003.parquet) apache-2.0 (dataset card: "This dataset is released under Apache 2.0"); its requests were written by Mistral-Small, DevStral-Small, GPT-OSS, Kimi-K2 and Qwen3-32B upstream, its tool descriptions are the MCP servers' own metadata 0df3cf37 test only (T5-tools and its copies), never trained

Rows and requests per split, counted on the training and validation records (*_train, *_val) and on t5_tools (the other files hold the same requests again):

source rows train rows val rows fresh (T5) requests train / val / fresh
glaive 6,567 281 — 1,880 / 80 / —
hermes 3,098 191 — 887 / 54 / —
toolace 25,362 1,381 — 7,245 / 394 / —
seal-tools 36,031 1,774 — 10,283 / 507 / —
api-bank 6,012 314 — 1,725 / 91 / —
when2call 13,885 847 — 5,832 / 356 / —
massive 50,989 8,899 — 14,559 / 2,541 / —
banking77 40,141 2,186 — 11,478 / 626 / —
bfcl-multiple — — — — / — / —
toucan — — 2,000 — / — / 2,000
all 182,085 15,873 2,000

The rows of each file are those the build wrote; how many requests each filter dropped is in the pre-registrations' amendments and in files/v5-slates.json, files/v5-plan.json and files/t5-counts.json (T5's counts; the full table is the repository's runs/s6-route7/counts.json). Of the teacher's rewrites, the slates show 12,999 of v4's and 822 of T5's.

How each file was checked

Before a row was read, every file was hashed and compared with the sha256 that the repository's run files record: runs/s6-route7/inputs6.json (the training and validation inputs of v7, amendment 1; files/inputs6.json) and runs/s6-route7/test-files.json with FRESH-OPEN (T5's five files, hashed before any system read them; files/FRESH-OPEN). The script refuses on any difference. Four kinds of file have no such record, and were checked otherwise (each file's check is in MANIFEST.json):

  • the blind copies (*_blind, *_blind_model): row by row, each is its parent with the request replaced by the placeholder, and the model copy is that file with each tool option's score 0.00 rank <position>; (the copies of v5_tools also against the hashes that sky-data route blind logged in the data folder, build7/logs7/blind-*.json);
  • the files read twice under two names (*.blind-bare.model.jsonl, the fix folders' copies of v5_tools): by sha256 against the published file;
  • the teacher's rewrites (teacher_rewrite_description_v4, _t5): every option that a record flags as rewritten, and every rewritten distractor of the probe copies, is a cut of its tool's rewrite in these files (or in v2's, skycmd-route's);
  • the files of skycmd-route read again: against the data folder's BUILD.json and FRESH.json.

Not in this repository:

  • build7/teacher/false-negatives.items.jsonl: the items of v5's false-negative filter (the training and validation hard slates before filtering); sky-data route v5-plan rebuilds them.
  • build7/teacher/false-negatives.jsonl: the teacher's 58,984 answers of v5's false-negative filter (option indices into the items above); their effect is in the hard slates of the records.
  • build7/teacher/false-negatives.refused.jsonl: items refused by the endpoint's moderation.
  • build7/fix/teacher/false-negatives.items.jsonl: the items of v6's fix (the rebuilt hard no-call slates of When2Call); sky-data route v5-fix-items rebuilds them.
  • build7/fix/teacher/false-negatives.jsonl: the teacher's 694 answers of v6's fix; their effect is in the hard slates.
  • build7/fix/teacher/false-negatives.refused.jsonl: items refused by the endpoint's moderation (empty).
  • build7/fix7/teacher/false-negatives.items.jsonl: the items of v7's fix (the rebuilt hard no-call slates without an anchor); sky-data route v5-fix-items rebuilds them.
  • build7/fix7/teacher/false-negatives.jsonl: the teacher's 296 answers of v7's fix; their effect is in the hard slates.
  • build7/fix7/teacher/false-negatives.refused.jsonl: items refused by the endpoint's moderation (empty).
  • build7/stage/plan5.jsonl: v5's plan of every slate (tool identifiers and cell levels); sky-data route v5-plan rebuilds it.
  • build7/fix/stage/plan5.jsonl: v6's plan; sky-data route v5-plan rebuilds it.
  • build7/fix7/stage/plan5.jsonl: v7's plan (--when2call-anchor nearest); sky-data route v5-plan rebuilds it.
  • build7/fix/stage/false-negatives5.items.jsonl: the false-negative items of v6's plan before the fix's job; sky-data route v5-plan rebuilds them.
  • build7/fix7/stage/false-negatives5.items.jsonl: the same for v7's plan.
  • toucan5/stage/rows.jsonl: the 26,134 Toucan rows that qualify (request, server, tools as Toucan gives them), the stage of T5's build; sky-data route toucan rebuilds it from the pinned SFT files (files/MANIFEST4.json); the 2,000 requests T5 uses are in t5_tools.
  • toucan5/stage/tools.jsonl: the stage's catalogue of the 1,770 Toucan tools (name, description, servers); the 1,448 that T5 shows are in teacher_rewrite_description_t5 with their name and description.
  • toucan5/stage/texts.jsonl: the text of each Toucan tool that the embedding reads (its name and description); rebuilt by sky-data route toucan.
  • toucan5/teacher/rewrite-description.items.jsonl: the items sent to the teacher (the Toucan tools' names and descriptions, given in original).
  • toucan5/teacher/rewrite-description.refused.jsonl: items refused by the endpoint's moderation (empty).
  • build4/teacher/rewrite-description.items.jsonl: the items sent to the teacher (the tools' own descriptions, given in original).
  • build4/teacher/rewrite-description.refused.jsonl: items refused by the endpoint's moderation (empty).
  • the embeddings (emb/), markers and logs of the build, the models and environment, the dumps and evaluation outputs (eval/), the model copies of T5 that the gate rendered (sky-data route render rebuilds them), and the raw upstream files (src/): scripts/build-route4-data.sh fetch downloads Toucan at the commit of files/MANIFEST4.json and checks every file's hash.

Limits

  • A failed model's data. These files trained a router that failed its gate. They are published so that the result can be checked, not as a recommended training set.
  • Known cues remain. v4 to v7 rebuilt the data against the cues that let a rule pick the right tool without reading the request; outside the slates v6 and v7 rebuilt, the records hold the cues that v5's data table names (docs/s6-route5-prereg.md, result block), and the anchor of v7's no-call slates is chosen with the request's own text (docs/s6-route7-prereg.md §10).
  • English mostly: 29,572 of the 199,958 counted rows are French, all from MASSIVE (requests in French, intent names and descriptions in English).
  • Synthetic upstream data: Toucan's requests and several tool-calling sets were written by language models upstream; the build filters and deduplicates, it does not check them by hand.
  • Personal data: no row holds text that its upstream set does not publish, apart from the teacher's rewrites of tool descriptions.
  • Keys: the builder masks key-like strings in requests and descriptions as <KEY> (sky_data::route::mask_keys). Every string of data/ and files/ was scanned with the same rule at assembly, and none is left: 53 matches are tool names (of ToolACE and Toucan) that were read by hand and are not keys, listed by hash in scripts/hf-publish-data.py (ROUTE_NOT_KEYS). The run files copied in files/ hold no absolute path of the build machine (the script refuses one).
from datasets import load_dataset
ds = load_dataset("maxgfr/skycmd-route7", "tools_train")