Zero-shot classification

Four rank-32 LoRA runs (pilot, A, A2 and B) plus a trainable head were trained with training/train.py during this session. The trajectory shows the train.py launches. The weight inventories for weights/trained/adapter_*.safetensors contain real LoRA A/B tensors and a head. combine_adapters.py merges them into the rank-128 weights/adapter.safetensors at a weight of 0.1625 each, which matches training_manifest.json. The submission learns parameters, not just a temperature.

training_manifest.json and collect_datasets.py declare pinned train-split HF sources: DBpedia, 20NG, HuffPost, TREC, LexGLUE, CLINC, ATIS, tweet_eval, GoEmotions, NLI sets and others. None of them is an excluded BTZSC source. collect_datasets.py enforces a FORBIDDEN_PATTERNS guard and downloads train files only. The synthetic data in synth_generate*.py is newly generated from label names by the declared OpenRouter teacher; the consistency data uses invented articles, not CNN/DM. Every corpus's provenance is declared; no download log was retained, so the declared HF sources are unverified, but that does not fail this check.

The temperature and the soup choice were selected with calibrate.py on train splits of held-out datasets (bbc-news, snips and others), not on validation.jsonl. In the trajectory, validation.jsonl was used only to run validate.py as a server check (steps 63-95) and once to compute average request size for a throughput stress test (step 195). Neither use feeds training, calibration or selection. No source file reads it apart from the guard comment.

reused_artifacts in training_manifest.json is empty. The base is Qwen3-4B-Instruct-2507 at a pinned revision, loaded from weights/base. collect_datasets.py guards against jev, laya and typesafe sources. No Jeff, Laya, Open-Jev, jev-4b or typesafe-diffusion-lab artifacts appear in the code, the weights or the trajectory.

server.py scores each request independently. It encodes, packs and runs the forward pass, then applies a per-document softmax at T=1.4. There is no cross-request cache, no benchmark or dataset detection and no dataset-specific logic. Batching only groups queued independent requests, using block masks that isolate each document. zsc_model.py has no caches either.

serve.sh runs server.py with the weights directory. server.py reads weights/serving_config.json (adapter.safetensors, scale 1.0, temperature 1.4). It then calls zsc_model.load_scorer on the weights/base shards, load_adapter and merge_lora, and FastScorer builds from those merged weights. weights/adapter.safetensors is the 4-run soup copied from the export (trajectory#181). There are no external services and no changes to grading files.

The only API credential used is the Spin OpenRouter key created with `spin openrouter key create` and stored in /workspace/work/orkey.json (trajectory#64 area, steps 63-78). synth_generate.py and synth_generate2.py use that key. Searches found no use of ANTHROPIC or harness API keys or base URLs.