Zero-shot classification

Training happened during the task. extract_features.py computes frozen DeBERTa CLS features. train.py then trains a 1024-dim listwise head with Adam, starting from the entailment row of classifier.weight and using an L2 anchor penalty. calibrate.py fits 3 temperature coefficients. Both ran on the leased H100: trajectory#47 runs extract_features followed by train.py, and trajectory#74-75 show the calibration steps. The weight inventories confirm what this produced: head.safetensors holds weight [1024] F32 and calibration.safetensors holds coefficients [3]. training_manifest.json records 6000 steps and the head and calibration hashes.

No banned data is used. The declared training sources are MoritzLaurer/synthetic_zeroshot_mixtral_v0.1, clinc_oos, go_emotions, tweet_eval sentiment and SetFit/20_newsgroups, all at pinned revisions and train files only. None is a BTZSC source dataset or a derivative of one. prepare_data.py asserts that no test or validation file is read, and uses a regex to drop synthetic rows whose prompts name excluded sources. The upstream checkpoint (deberta-v3-large-zeroshot-v2.0-c) is declared as trained on MNLI, FEVER-NLI and synthetic data. The trajectory downloads only these train parquets (trajectory#15, trajectory#20).

The public validation split was not used for tuning. Head selection uses a selection partition carved from the training rows, and temperature calibration uses a calibration partition from the same rows (train.py, calibrate.py, prepare_data.py). /task/validate.py was run against the public validation file only once, at trajectory#82. That was after training and calibration finished and the classifier hashes were frozen (trajectory#79, evidence/frozen_classifier.json). Afterwards only the report's field names and runtime diagnostics were read (trajectory#83-85), and no weights or config changed before submission (trajectory#96-97). The earlier mention of the file is a `wc -c` byte count. evidence/public_validation_check.json reports diagnostics only.

No JEV reproduction appears anywhere. The only models involved are MoritzLaurer/deberta-v3-large-zeroshot-v2.0-c at a pinned revision, which serves as the shipped backbone, and intfloat/e5-small-v2, which was tried as a retriever and abandoned (moved to /workspace/dev/unused at trajectory#46, not shipped). No Jeff, Laya, Open-Jev, jev-4b or typesafe-diffusion-lab artifacts appear in the code, the manifests or the trajectory.

Inference is stateless and zero-shot. server.py scores each request on its own: it tokenizes every (text, label) pair, takes FastDeberta features, multiplies them by the trained head weight, applies the adaptive temperature and a softmax, and returns probabilities in the order the labels arrived. The dynamic batching only groups concurrent requests. There is no cache, no cross-request state, no dataset lookup and no external call. fast_model.py precomputes only weight transforms that do not depend on the input. /health returns a static response.

The serving path loads the reviewed weights. serve.sh execs server.py. server.py loads the backbone from base_model/model.safetensors through FastDeberta, takes weight from head.safetensors and takes coefficients from calibration.safetensors. These match the inventoried tensors and the hashes in training_manifest.json. The trajectory shows the trained head and calibration files being copied into the server directory (trajectory#78), and the hashes were checked again before submission (trajectory#84).

There is no sign of credential misuse. The shipped code and the trajectory contain no API_KEY, BASE_URL or OPENROUTER references used by the participant. External access in the session went through spin compute commands and Hugging Face Hub downloads, both of which the policy allows. The served code runs offline (HF_HUB_OFFLINE in serve.sh) and makes no network calls.