ekVachan-decoder

LoRA adapters (r=16, alpha=32, 128 target modules) for Qwen/Qwen3.5-4B, the decoder arm of ekVachan -- an open, self-hostable alternative to TypeSafe AI's Jev System-One decision model. Given a state and a set of typed questions (choice, score, noul), it returns calibrated probabilities in one non-autoregressive forward pass (a restricted-logit read over single-token letter codes), not a generated response.

Full source, training pipeline, and evidence trail: github.com/asp616848/better-jev-for-all (README.md for the summary, PRD.md for every measured number and how it was produced).

Need speed over accuracy? Try the 0.8B sibling — abhi6168/ekvachan-decoder-flash: same API, ~24ms typical latency, trained through the same curriculum.

What's in this repo

Two named LoRA adapters over the same base model, hot-swapped at serving time by serve.inference.RoutingDecoderModel:

• ekvachan-decoder-qwen-benchcorpus/ -- text-only, max_options=26
• ekvachan-decoder-qwen-vision/ -- the strongest (stage3) checkpoint, vision-capable, max_options=588; also answers text-only requests by default in this project's reference server (EKVACHAN_TEXT_ADAPTER=vision)

Each subfolder carries its own manifest.json (training config, real measured accuracy/Brier/ECE) alongside the adapter weights -- treat that file as the source of truth over this card if they ever disagree.

Real, verified numbers

Third-party benchmarks, zero training exposure, measured against the actual serving class (RoutingDecoderModel, eager and CUDA-graph modes identical within noise -- see PRD.md 13a.33):

Benchmark	Accuracy
JevBench (231/231 items)	70.99%
jabr-v2 (944/944 items)	85.49%

Latency, on our own dedicated reference server (CUDA graphs on by default, PRD.md 13a.34): ~112ms p50 / ~131ms p95 on realistic (100-250 token) request shapes; 45ms on short (15-token) synthetic prompts. This model is not yet competitive with the fastest field entries (Von <18ms, Laya ~16ms) on raw latency -- see the main repo's README for the full comparison and an honest accounting of what CUDA graphs did and didn't fix.

Usage

Not a plain AutoModelForCausalLM.generate() model -- it requires the restricted-logit read mechanism (build a single-letter code-table prompt, take the last token's logits, restrict to the option codes, softmax). Use serve.inference.RoutingDecoderModel from the main repo directly rather than reimplementing this:

from serve.inference import RoutingDecoderModel

model = RoutingDecoderModel(checkpoints_dir="path/to/this/repo's/local/clone")
result = model.predict_choice(
    "The customer says the package never arrived.",
    ["refund", "replace", "escalate"],
)
print(result["choice"], result["probabilities"])

Or try it with zero setup in the ekVachan Space (note: Space latency includes ZeroGPU queueing, not representative of the numbers above).

License

Apache-2.0, same as the base model and the main repo.