Loading…
Loading…
32B is trained, its spoke is vox-lang and its gate is not evaluated: the records that decide 32B are missing: collateral_damage_report.json, gate_receipt.json.
The largest rung, planned for machines at the top of the memory range and not available yet. It is meant for the hardest cross-file work once a smaller size shows the approach holds. Until it ships, pick a smaller size. The measured training peak for 32B is 353743 MiB (peak_source: run_summary), and no serving memory figure has been measured for it.
No held-out report is on record for 32B, so no example is shown for it.
The 32B adapter sits on the base pin Qwen/Qwen3-32B@9216db5781bf21249d130ec9da846c4624c16137, and this command downloads it after checking the BLAKE3 hash of every file in hf_sync_manifest.json.
vox mens hub pull --size 32B --namespace brbrainerd --allow-ungated
vox mens serve loads the 32B adapter only from a run directory that holds a passing collateral report and a passing gate_receipt.json, so the pull needs --allow-ungated and no serving example is given.
The docs page explains why the 32B adapter cannot be loaded with transformers or PEFT.
| Result for 32B | Held-out pass@1 | Collateral accuracy | McNemar p-value |
|---|---|---|---|
| Base 32B | not recorded | — | — |
| Adapter 32B | — | — | — |
The 32B rows compare the base model with the adapter on identical items. No held-out report under the current prompt protocol qwen3_non_thinking_codegen_v1 exists for 32B, so no held-out score is shown for it.
32B ran 168 optimizer steps over 1185861 tokens, ended with a final loss (EMA) of not recorded, a mean train loss of 0.6644 and a last-epoch validation loss of 0.5461, took not recorded and peaked at 353743 MiB, from Vox commit b2ce949b4; the settings every size shares are on the docs page.
The 32B repository holds candle_qlora_adapter.safetensors, the adapter, training and gate manifests and records, evidence/, and hf_sync_manifest.json with the BLAKE3 hash of every file.
This card is generated by vox mens hub sync; the ladder index lists the other sizes and how the cards are made is on the docs page.
Qwen/Qwen3-32B