arrochi112/OpenGrad-ToolPolicy-Canonical-v2-minus-xlam

This is the byte-identical training view for a joint xLAM-plus-CALL_PREDICTION removal experiment. xLAM is currently the corpus's only source of that supervision contract, so this is not a pure source-content ablation. It carries no result and is not a recommended mixture.

What this is

OpenGrad-ToolPolicy-Canonical-v2 with one source removed: xLAM/APIGen. Three sources remain, 115,895 canonical records, 118 shards.

It exists because OpenGrad planned paired fixed-compute and matched-exposure runs against this selection, and an experiment is only checkable if its input is available. This view is their exact input. Removing xLAM also removes all CALL_PREDICTION supervision, so the runs estimate that joint intervention rather than an xLAM-content-only effect.

It is a selection, not a rebuild

The parent corpus writes its shards per source, so removing a source is the removal of that source's shards and nothing else. Every retained shard here is a byte-for-byte copy of the parent's: no record was parsed, rewritten, re-encoded, or re-sharded.

That is verifiable rather than asserted. Every shard's SHA-256 is recorded in release-manifest.json under output_shards, and each one equals the parent's hash for the same file. The manifest also records, under derived_view, the parent's fingerprint, its Hub revision, and the names and hashes of the 58 shards that were excluded — so the exclusion can be checked against the parent dataset as well as against this one.

parent fingerprint : 8ced403b996e563d6e279aee7fdb346fc829fe5ff6af9daf8ef47c0a4007e161
current fingerprint: 5fc739040b1a09fa7b8ac20494609bd4cbe89efc6ed54e4dbc9c2df12268b7f3
prior fingerprint  : f8ba687e16d8ab740b78a503adb41f4be28bd775f925291d6da7ec31be1ac5ac
                     superseded only because its interpretation metadata called this a source
                     ablation without stating the inseparable supervision-channel removal;
                     retained parquet bytes and counts are unchanged

Consequences worth stating: supervision contracts, renderer behaviour, loss masks, contamination decisions, source provenance, and trainability decisions are identical to the parent for every retained record. Mechanically, the row-selection difference is that xLAM is absent. Experimentally, that same removal eliminates the entire CALL_PREDICTION channel because no other source currently provides it.

What was removed, and what it costs

Source	Canonical records	Trainable	Shards
glaive-function-calling-v2	98,339	97,112	99
toolace	11,051	2,259	12
when2call	6,505	6,505	7
retained total	115,895	105,876	118
xLAM/APIGen (removed)	57,342	56,090	58

xLAM is 34.6% of the parent's trainable records but only 11.7% of its loss-bearing tokens. Those two numbers differ because xLAM records are short: under its CALL_PREDICTION contract the supervised target is the tool call itself and no tool result or final response follows it, so a record ends where a COMPLETE_TRAJECTORY record would continue. Measured across the parent's rendered samples:

	full corpus	minus xLAM	ratio
trainable records	161,966	105,876	0.654
rendered tokens	99,717,548	64,146,939	0.643
supervised (loss-bearing) tokens	33,565,721	29,630,369	0.883

That asymmetry is the reason the ablation needed two arms and two different step counts rather than one. It is also the most transferable thing this view demonstrates: in a heterogeneous corpus, a source's share of records is a poor proxy for its share of training signal, and an ablation that matched on record counts would have confounded two variables.

Supervision contracts

xLAM was the only source carrying CALL_PREDICTION, so this view is entirely COMPLETE_TRAJECTORY — 105,876 trainable records. The configs explicitly declare that complete post-filter contract set; readiness rejects any retained empty or absent selection. Although the implementation filters by source, the experimental treatment necessarily removes both xLAM and the supervision channel.

Contract	Trainable here	In the parent
COMPLETE_TRAJECTORY	105,876	105,876
CALL_PREDICTION	0	56,090

Intended use

As the input of the paired joint xLAM-plus-CALL_PREDICTION removal experiment, and as a comparison artifact for anyone examining the parent corpus. Training on it is not a recommendation: the three remaining sources are the parent's, unbalanced and unmodified, and the parent was never a recommended mixture either. No xLAM-specific causal attribution is valid unless another evidence- backed CALL_PREDICTION source separates source identity from supervision type.

Not suitable as a benchmark. A subset of a training corpus is still training data.

Attribution

The parent's CITATIONS.bib and source-licenses.md are inherited verbatim, so they still list xLAM alongside the three retained sources. That is deliberate: attribution files should over-include rather than risk dropping a licence obligation. xLAM's entry documents the source this view removes, and the parent dataset is where those terms continue to apply.