Tern-35B-A3B

Tern is a typed decision model: supply a state and named questions with their complete answer options, and receive a probability distribution for every question. It uses a Qwen3.6-35B-A3B backbone with a jointly trained decision head. It does not generate reasoning or text tokens.

This release is optimizer checkpoint 1674, trained on 499,999,909 tokens. It contains merged training-rounded BF16 backbone/adapter weights and a trained head stored in FP32. The frozen serving path precasts selected head parameters to BF16 and uses BF16 autocast. All 40 layers and eight routed experts are active. No early exit, expert pruning, external tools, retrieval or test-time augmentation is used.

Use the tested Tern Docker serving repository for RTX PRO 6000, A100 and H100 profiles. The default RTX BF16 profile includes the cuBLAS workspace reproducibility setting automatically; no manual patch is required. The model, results and serving code are public; no access token is required. The original frozen A100 runtime is retained for provenance; the dedicated repository is the RTX submission entrypoint. The optional H100 FP8 profile is a separate experiment. A stock Transformers causal-language-model loader does not implement the trained decision head. Capacity is 32,768 encoded tokens for the entire state and all questions/options; oversized requests are rejected without truncation.

Public evaluation

Complete public 0.3 run	Public index	Raw index	Requests
RTX PRO 6000, fixed-workspace Docker	59.01	68.93	140,178
Original A100	59.10	69.04	140,178

The fresh RTX run includes all 37 index and six auxiliary benchmarks: 140,176 successful requests, 2 unsupported BRIGHT requests, and zero errors. The official scorer reports complete: true, and an independent local rescore reproduces every score field. All request IDs and payload hashes match the frozen suite. Complete predictions, exact settings, per-benchmark metrics and comparisons are in the results repository, under runs/step1674-rtx-top8/. The older A100 run remains unchanged under runs/step1674-top8/.

These are public scores, not the private Full score or an accepted leaderboard placement. See TRAINING_DISCLOSURE.md for public development exposure and training-overlap candidates; this is not an uncontaminated held-out claim.

Reproducibility and limitations

The RTX launcher pins CUBLAS_WORKSPACE_CONFIG=:4096:8. The fixed serial panel matched across four fresh processes; head tensors and numerical source files were unchanged. This mitigates observed restart variation on that panel, not a universal determinism guarantee. Cross-GPU and cross-batch probability equivalence remain unproven. In the matched full public runs, RTX versus A100 changed 4,384/312,256 selected decisions (1.404%). Both hardware and batch scheduling differ; these effects are not separately isolated. No numerical tolerance was widened.

A separately warmed 128-request RTX serial panel after the full evaluation measured median 59.48 ms, mean 74.09 ms, p80 89.38 ms and p95 134.59 ms. These include HTTP, encoding and inference, and exclude model loading and warmup. The concurrent full-evaluation timings include queueing and are not single-request latency. Maintainers must still run their private quality and latency checks.

The model is an experimental decision scorer. Its probabilities are not guarantees of correctness. It is not a generative assistant or multimodal release.

Provenance and attribution

Base revision: Qwen/Qwen3.6-35B-A3B@995ad96eacd98c81ed38be0c5b274b04031597b0. Tern model weights, the trained head and bundled Tern code are licensed under Apache-2.0. The base model's original license is preserved in BASE_MODEL_LICENSE, and the CLEF-derived runtime retains its original notice. See NOTICE and TRAINING_DISCLOSURE.md for attribution and training-source evidence. The license choice does not rewrite source dataset terms.

export.json records the checkpoint and training/validation fingerprints. model-file-sha256.json records the inference-file bytes. Optimizer state, training data, benchmark inputs and private traces are not included.

Upstream SGD auxiliary-input correction

A final upstream check found commit 2d7791c, which fixes an SGD builder bug: rows previously retained a reference to the growing dialogue history and therefore included future turns. All 2,500 SGD requests in both frozen full runs are affected. Their original SGD scores should not be interpreted as valid current-turn intent classification results.

A separate supplemental check keeps the same selected request IDs, schemas, labels, model and head, but stops each history at its recorded current user turn, matching the corrected builder's snapshot semantics. All 2,500 requests succeeded. Corrected SGD macro-F1 is 0.7033, versus 0.1052 in the frozen RTX run (original A100: 0.1071). Independent local rescoring reproduced the corrected benchmark summary exactly. Supplemental results include provenance and separate predictions.

SGD is excluded from the headline index: 59.01 public / 68.93 raw remains unchanged. The complete canonical frozen runs and their payload hashes are preserved untouched. The supplemental corrected requests have different payload hashes and are not silently substituted into those runs. Upstream HEAD checked was f5a6e4fd85fed05e77878dfc4966d0568bfb45ec; its scoring/edition files and published suite hash declarations are unchanged from the frozen scorer pin. This discrepancy is disclosed for maintainer review. No benchmark input text is redistributed.

Submission status

The model, results and serving code are public. Decision Index submission PR #227 is open for review. Inclusion and the maintainers’ private quality and latency checks are pending. The evaluated inference files and complete predictions are unchanged.