krishna765/deck-4b-v1.0

deck-4B v1.0

Rank-8 LoRA for calibrated decision probabilities. The served model id is deck-4b-v1.0.

We trained this adapter on our data. We did not train the base checkpoint. The base checkpoint name, revision, and license are in adapter_config.json (base_model_name_or_path) and in the code repository README. Leave that field unchanged; the adapter will not load against a different base.

Configuration

joint_config.json holds the calibration temperature used at inference:

{"temperature": 1.159999966621399, "readout": "restricted A-P next-token logits"}

It was fitted by NLL on this adapter's 65-row validation split.

Training

• One bf16 epoch, learning rate 1e-5, LoRA rank 8, alpha 16, dropout 0.05 over Qwen3.5 attention and linear-attention projections.
• First 500 records of a seeded, nested candidate mix; a stable hash split gave 435 training and 65 validation rows.
• The mix contains synthetic tool selection, SNIPS and smart-intent examples, GLiClass/GliClass-RAC records, redistributable Open-Jev synthetic/control records, and deterministic RLCD examples.

No JevBench item was used for gradient training or temperature fitting. The 231 public JevBench items were evaluated during development to select this adapter and its FP8 serving configuration, so public scores are development-directed. An exact normalized-state and shared eight-word-sequence audit of the full 1,000-row candidate pool against all 231 public items found zero hits.

Serving

FP8 weight-only serving for the JevBench TypeSafe protocol needs an FP8-capable NVIDIA GPU (tested on one H100 80 GB). Install and start from the code repository README:

phasor-jevbench-fp8 \
  --adapter krishna765/deck-4b-v1.0 \
  --name deck-4b-v1.0 \
  --host 0.0.0.0 --port 8090

The server exposes POST /v1/systemone and GET /health. It supports noul, choice, and score, and returns native probabilities.

Local JevBench public reference

JevBench revision 1bcc55eb6c8cffde2306b3db03ede39b61c6152a, official typesafe adapter, loopback H100, one request at a time, FP8 runtime:

Tier	Accuracy	ECE10	p50	p95
Easy	48/48	0.017	50.0 ms	50.5 ms
Original	68/72	0.051	50.0 ms	50.6 ms
Hard	87/111	0.041	60.5 ms	141.8 ms
Overall	203/231	0.023	50.6 ms	116.9 ms

These are not official leaderboard numbers.

Limits

• English only.
• Up to 16 options in one pass; larger sets use knockout passes.
• Prompts above 16,384 tokens are rejected, never truncated.
• FP8 numerics are hardware- and runtime-sensitive.

License

Apache-2.0.