Datasets:
Submission: yotambr/safedecide-140m (general, then fine-tuned on train): accuracy 0.754
Hi @codelion , I'd like to add SafeDecide-140M to the fitted-on-train table.
| Field | Value |
|---|---|
| Model | yotambr/safedecide-140m, 140M, mmBERT-small cross-encoder, Apache-2.0 |
| Weights revision | fcd96ad68b2e05cbc59ad2697919dfc6bda616f1 |
| Kind | general, then fine-tuned on train |
| Test split | all/test: 400 cases, 2,000 decisions, all answered |
| Inference | one encoder pass per option, no generated tokens. Each question is scored on its own, not all five in one request |
| Latency | 80 ms p50 for a 3-option decision, fp32 ONNX on one CPU thread (Colab VM) |
| Model | Kind | Accuracy β | KL from gold β | Brier β | ECE β | Reported latency |
|---|---|---|---|---|---|---|
| SafeDecide-140M (mmBERT-small, 140M) | general, then fine-tuned on train |
0.754 | 0.126 | 0.069 | 0.147 | 80 ms, 1 CPU thread |
On the definitions: KL and Brier are per decision against the soft gold, and my scorer reproduces your Uniform row exactly (0.444 / 0.238). Accuracy is agreement with the gold label. If a tie in the gold counts as correct, as in score_model.py in SaylorTwift's evals bucket, it's 0.760. ECE is 15 equal-width bins on top-option confidence, pooled over the 2,000 decisions. With 10 bins it's 0.154, and per question (10 bins) averaged over the 20 questions it's 0.169. Tell me which convention you want in the row and I'll match it.
By question type
| n | Accuracy | KL | Brier | ECE | |
|---|---|---|---|---|---|
| noul | 600 | 0.840 | 0.065 | 0.049 | 0.111 |
| choice | 600 | 0.717 | 0.167 | 0.084 | 0.158 |
| score | 800 | 0.718 | 0.141 | 0.072 | 0.176 |
By workflow
| n | Accuracy | KL | Brier | ECE | |
|---|---|---|---|---|---|
| agent_trace_observability | 500 | 0.738 | 0.083 | 0.049 | 0.187 |
| customer_service | 500 | 0.730 | 0.190 | 0.099 | 0.137 |
| invoice_processing | 500 | 0.792 | 0.154 | 0.083 | 0.085 |
| security_incidents | 500 | 0.756 | 0.078 | 0.044 | 0.189 |
Training. I split train by case into train / validation / calibration (80 / 10 / 10). Stage 1 trains on the openjev_v2 stream of jev-distill-corpus-v3 plus the train part. Stage 2 fine-tunes on the train part alone. The checkpoint was picked by validation NLL, the per-type temperatures were fitted on validation, and the automate/escalate thresholds on calibration. Nothing from test was used: no rows, no labels, nothing derived from predictions on it. I also checked the outside data: none of the 400 test states, and none of the long text fields inside them, appear anywhere in the jev-distill train, validation or calibration splits I used (19,389 distinct states).
Predictions. Every test decision with its full distribution is in yotambr/safedecide-results, viewer tab typed_decisions_test, model SafeDecide (no ConfTr), so you can run your own scorer on it. The same numbers are in the model repo's .eval_results/typed-decisions.yaml. Code and notebooks: github.com/yotambraun/safedecide.
Something I measured along the way. The model ships with an automate/escalate threshold certified on held-out data (Learn-then-Test with exact binomial bounds). I ran the same calibration on every model in SaylorTwift's per-decision results, to see what share of the test decisions each one could take alone with the error among them certified at most 5% (100 random case splits, Ξ΄ = 0.1). The order is quite different from accuracy: JEV-27B-VL gets 20.4% with an accuracy of 0.688, d3 gets 0.6% with 0.749. Table: automation_board.md, script: scripts/automation_board.py. If a column like that would help the card, I can open a PR.
Thanks for running the benchmark.
Thanks for the detailed report and the public predictions. SafeDecide-140M is now on the card in the fitted / fine-tuned table at #17: 0.754 accuracy, KL 0.126, Brier 0.069, and ECE 0.147 marked as your 15-bin definition, with 80 ms on one CPU thread as reported. The notes on the rows mention that it scores each question on its own rather than all five in one request.
SafeDecide v0.2 is fixed before scoring test: SHA-256 11a60a13db03e1e8c831d7003d7a51bb8e20c4acc2a15b8d6a013872251812e7 of recipe.json, which records the recipe (chosen by 5-fold cross-validation on train only) and the weight hashes of two models, SafeDecide-140M v0.2 and SafeDecide-400M v0.2. I'll score test once and report both results here, whatever they are, together with recipe.json.
Hi @codelion , here are the results for the two sealed models (seal above: 11a60a13β¦812e7). test was scored once, after the seal was posted. I'd like to add SafeDecide-400M to the fitted / fine-tuned table. The 140M v0.2 did not beat v0.1, so please keep that row as it is.
| Field | Value |
|---|---|
| Model | yotambr/safedecide-400m, 400M, Ettin-encoder-400m cross-encoder, Apache-2.0 |
| Weights revision | 229219211f24674ff8c2dc6a7f1e389c82704892 |
| Kind | fine-tuned on train, no other decision data |
| Test split | all/test: 400 cases, 2,000 decisions, all answered |
| Inference | one encoder pass per option, no generated tokens. Each question is scored on its own, not all five in one request |
| Latency | 605 ms p50 for a 3-option decision, PyTorch fp32 on one CPU thread (my laptop; the 140M takes 80 ms there too) |
| Model | Kind | Accuracy β | KL from gold β | Brier β | ECE β | Reported latency |
|---|---|---|---|---|---|---|
| SafeDecide-400M (Ettin-encoder-400m, 400M) | fine-tuned on train |
0.787 | 0.083 | 0.045 | 0.171 | 605 ms, 1 CPU thread |
SafeDecide-140M v0.2 (mmBERT-small, 140M; branch v0.2) |
general, then fine-tuned on train |
0.748 | 0.113 | 0.061 | 0.140 | 80 ms, 1 CPU thread |
| SafeDecide-140M v0.1, the current row | general, then fine-tuned on train |
0.754 | 0.126 | 0.069 | 0.147 | 80 ms, 1 CPU thread |
Paired against v0.1 on the same decisions (95% bootstrap interval over cases): the 400M is +0.033 [+0.014, +0.052], the 140M v0.2 is β0.006 [β0.026, +0.011].
Same conventions as my first post: accuracy is agreement with the gold label (counting a tie in the gold as correct: 0.794 for the 400M; bench/score_with_jobs.py, which takes the first tied option, gives 0.784, and 0.750 for v0.1). ECE is 15 equal-width bins pooled over the 2,000 decisions; with 10 bins it's 0.172, and per question (10 bins) averaged over the 20 questions it's 0.181. These numbers are the sealed pass, on an A100 with bf16 autocast. In full precision on CPU (what the thresholds were certified on) the same weights give 0.7875 / 0.0825 / 0.0455 / 0.172 for the 400M, 0.755 / 0.113 / 0.061 / 0.147 for the 140M v0.2 and 0.7545 / 0.126 / 0.069 / 0.148 for v0.1, and the 400M's paired gain is the same, +0.033 [+0.014, +0.052]. Those predictions are in the results dataset's v0.2/full_precision/.
By question type (400M)
| n | Accuracy | KL | Brier | ECE | |
|---|---|---|---|---|---|
| noul | 600 | 0.858 | 0.042 | 0.033 | 0.135 |
| choice | 600 | 0.748 | 0.115 | 0.060 | 0.169 |
| score | 800 | 0.762 | 0.089 | 0.045 | 0.202 |
By workflow (400M)
| n | Accuracy | KL | Brier | ECE | |
|---|---|---|---|---|---|
| agent_trace_observability | 500 | 0.706 | 0.089 | 0.053 | 0.157 |
| customer_service | 500 | 0.822 | 0.094 | 0.047 | 0.196 |
| invoice_processing | 500 | 0.842 | 0.078 | 0.044 | 0.129 |
| security_incidents | 500 | 0.778 | 0.069 | 0.038 | 0.209 |
How the choices were made. All on train. I chose the recipe by 5-fold cross-validation of the 140M model on the 1,080 train cases outside the same 600-decision calibration carve as v0.1. Out-of-fold accuracy was 0.658 for v0.1's schedule and 0.748 for the chosen recipe (longer schedules with early stopping, a longer first stage, and training-only questions built from the factors of train cases). The 400M uses that recipe, fixed in advance. It is trained straight from the base encoder on 972 train cases, with early stopping on the other 108, and has no other decision data. The thresholds are certified on the calibration carve. recipe.json records the code commit, every out-of-fold score, the choice, and the weight and config hashes of both models; its SHA-256 is the seal. One disclosure: before sealing I compared both models with v0.1 on the calibration carve (+0.048 for the 400M there). test was not used for anything before the scoring pass.
Predictions. Every test decision with its full distribution is in yotambr/safedecide-results, tab typed_decisions_test, models SafeDecide-400M v0.2 and SafeDecide-140M v0.2. The 400M's numbers are also in its repo's .eval_results/typed-decisions.yaml.
The automate/escalate side. At a certified error of at most 10% (100 random case splits of test, Ξ΄ = 0.1), the 400M can take 52.4% of the test decisions alone, against 33.2% for v0.1. Its shipped threshold, applied unchanged, automated 39.7% of them, and 4.3% of those were wrong.
Thanks again for maintaining the card.
