fly88oj/decidex-core-8b

Decidex v7 — the flagship adapter (84/86 official agreement)

LoRA adapter (r=32) for Qwen/Qwen3-8B that turns a frozen chat model into a Jev-style decision engine: state in, typed decisions with probabilities out, one forward pass, zero generated tokens.

Distilled from 17,954 samples of the official Jev API's actual outputs (four generation + active-mining rounds; corpus and pipeline in the Decidex repo, see REPRODUCE.md). This is the only public model lineage trained toward the official model's real answers rather than synthetic labels.

Measured agreement with the official API

On the 87-question comparison corpus (official answers collected live):

Primitive	Result
Choice top-1	23/23 (100%), distribution JS divergence 0.0025
Noul decisions	51/52 (98.1%)
Score modal level	10/11 (0.909)
Overall	84/86 (97.7%)

Generation-distribution disagreement vs the official model: 4.6% (mining-round measurement; the 4B baseline was 26%).

Usage

Serve through the Decidex service (byte-compatible with the official API — both official SDKs verified):

pip install -e '.[all]'
decidex serve --engine llm --model Qwen/Qwen3-8B \
  --lora <this-adapter> --device cuda:0

Or use GGUF builds of this adapter (merged) for llama.cpp / Ollama / LM Studio — see the decidex-gguf repo.

Training

• Base: Qwen/Qwen3-8B (frozen, BF16)
• LoRA r=32 α=64 on q/v/o_proj — 0.24% trainable
• Soft-label cross-entropy on the letter-logit readout position
• Dataset: distill_dataset_v4.jsonl (7,249 samples = broad sweep + score-heavy + first active-mining round ×2), 2 epochs, bs 4 × accum 2
• Hardware: one RTX 4090 24GB, ~80 min

Honest notes

• Probabilities are distilled, not RLCD-trained; residual disagreements with the official model (~2-11% depending on primitive) are semantic and documented in COMPARISON.md of the repo.
• Not affiliated with TypeSafe AI; Jev and TypeSafe are their trademarks.