OpenGrad-Qwen3.5-2B-M0-ABL-MinusXLAM-MatchedExposure

OpenGrad — Qwen3.5-2B, joint xLAM + CALL_PREDICTION removal (matched-exposure)

Full-parameter SFT of Qwen/Qwen3.5-2B on Canonical-v2 with xLAM removed, training for a horizon computed to match the reference's supervised-token exposure (2,119 steps; the logged exposure came out 1.19×, see below). Produced by OpenGrad.

Research artifact, not a production model. Published because the experiment's whole point is a negative result that is only checkable if the weights behind it are available.

What this experiment actually removed

Canonical-v2 maps xLAM to every CALL_PREDICTION record and the other three sources to COMPLETE_TRAJECTORY. Removing xLAM therefore removed a source and the corpus's entire call-prediction supervision channel at the same time:

this model measures the effect of removing xLAM together with the CALL_PREDICTION channel.

It is not a pure xLAM-content ablation, and no xLAM-specific causal claim can be based on it. Source identity and supervision type are perfectly aligned in this corpus.

The step count is computed, not chosen

matched_steps = 2400 x (29,630,369 / 33,565,721) = 2118.6 -> 2119

Matching on supervised (loss-bearing) tokens rather than records or rendered tokens, because that is the mass the loss acts on. xLAM's records are short, so the measures disagree sharply (records → 1,569, rendered tokens → 1,544, supervised tokens → 2,119). This arm additionally spends 11.7% fewer optimizer steps/FLOPs than the reference.

The match holds for the corpus totals, not for what training saw. Batches are token-budgeted, so steps do not hold supervised tokens fixed: logged, this arm saw 6,743,788 supervised tokens against the reference's 5,678,531 (1.19×, not ~1.0×).

Results (confirmatory partition, 1,277 examples, scored once)

Recall collapses furthest here while precision is highest and over-calling is lowest — the same direction as the fixed-compute arm, run further. The arm that trains more does better, which is consistent with the recall loss tracking the missing supervision, but exposure is not ruled out: the metrics above come from checkpoints that saw fewer supervised tokens than the reference's selected checkpoint (3.36M here at 1060 and 3.80M for the fixed arm at 1200, against 4.27M for the reference at 1800). "The reduced budget does not explain the result" is not supported as stated. The two arms are one experiment and must be read together.

Honest limits

Not a promotion. Every checkpoint of both arms is REJECTED on regression.call_recall against B0 (every matched-arm DEV checkpoint and fixed-arm 1800/2400 also fail call_f1_retention). The gate was left exactly as written.

Single seed. A small between-arm difference is a finding to replicate, not a settled result.

Provenance. The matched-exposure arm was launched at 12:49 UTC on 2026-09-11 from commit 500cb4e with uncommitted changes (git_dirty: true in experiment.json and events.jsonl). 0807fa3 is the 13:26 UTC commit that recorded the finished run, not the launch commit.

Some behaviours are unmeasured. Tool-selection accuracy, argument validity and schema validity are not computed; their absence is not a zero.

Checkpoints

All four retained checkpoints are published under their step: checkpoint-530, checkpoint-1060 (the DEV-selected one), checkpoint-1590, checkpoint-2119. Selection used the 2,373-example DEV partition and the rule frozen before the run; the confirmatory partition was scored once, on checkpoint-1060 only.

Intended use

Research artifact. Not safety-tuned, not aligned, not for autonomous tool use. It inherits the limitations of its public training sources, including synthetic data and unverified tool invocations, and it will behave poorly outside the decision-boundary behaviour it was trained for.

License and attribution

These weights are derived from Qwen/Qwen3.5-2B (revision 15852e8c), released under the Apache License 2.0 (license text at that revision). That license applies to this derivative: keep the license and its notices, and note that these weights are modified from the original by the post-training described above.

The training data are modified derivatives of upstream datasets with their own terms (CC-BY-4.0 and Apache-2.0, attribution required), listed per source in the dataset card's source-licenses.md. license: other / composite-per-source records that no single license covers every component; it does not relicense any of them.