Solar-Decision-JEV-27B-V2

The r32 + solar(V2) merge. huihui Qwen3.8-27B-abliterated + the AutoTrust r32 lm_head LoRA + our solar r16 LoRA re-trained to step 1894 (second full pass over the corpus), merged by direct LoRA-delta arithmetic.

This is the V2 of SargeDev/Solar-Decision-JEV-27B. Same architecture and merge recipe; the only change is the solar half, trained far past the V1 checkpoint (step 1219 → step 1894, ~55% further through the cosine schedule with a full second epoch of data).

The decision readout is trained INTO the model, so a typed decision is a single prefill with max_tokens=1, constrained to the option tokens, read back as log-probabilities. No separate head to run. (vLLM only — llama.cpp does not return the per-token logprobs this path needs.)

Gate results, honestly (27,695 rows, same harness/head/calibration as every number below)

What training solar harder bought us: the combined merge improved on V1 on every cell (choice +1.2 pts, score +2.3 pts, AUROC +0.002, overall +2.2 pts). The direction is real. But it is still below r32-alone on the decision cells, for the same structural reason: r32 trains the decision distribution into lm_head; solar has no lm_head LoRA, so stacking its delta dilutes a readout r32 had already specialized. More solar training shrinks the penalty; it does not erase it.

One calibration caveat, stated plainly: choice ECE here is 0.145 (r32-alone is ~0.03 on this kind). V2's typed-decision top picks are decent; its typed confidences on choice are less trustworthy than r32's. noul ECE is fine (0.007).

If you want the best decision instrument we ship, use SargeDev/Solar-JEV-27B-R32 (r32 alone) or SargeDev/JEV-27B-Uncensored. Use this V2 if you want the merge of both lines (e.g. for the everyday-chat behavior the solar line carries) and accept the measured cost above.

Everyday chat

Same as the rest of the Solar family: plain-English conversation is the solar line's home turf — crisp calls with honestly stated confidence, no hedge-fest. Pass enable_thinking: false in chat_template_kwargs (or --default-chat-template-kwargs at serve time) or the template emits its reasoning trace into the answer.

Serving

vllm serve SargeDev/Solar-Decision-JEV-27B-V2 \
  --served-model-name solar-jev-27b-v2 \
  --trust-remote-code --max-model-len 8192 \
  --default-chat-template-kwargs '{"enable_thinking": false}'

Quantizations of this model

• SargeDev/Solar-Decision-JEV-27B-V2-GGUF — 11-quant ladder (IQ4_XS…Q8_0), nextn/MTP drafter tensors preserved for speculative decoding. Fast everyday chat; not a typed-decision instrument (llama.cpp exposes no token logprobs).
• SargeDev/Solar-Decision-JEV-27B-V2-NVFP4-FP8 — NVFP4 W4A4 (attn + gate/up) + FP8 down_proj via llm-compressor, serves in vLLM with --quantization compressed-tensors.

Lineage & credits

• AutoTrust AI — the r16/r32 decision LoRA recipe, the 24-slot decision head, and the serve/decision machinery (autotrust/JEV-27B).
• huihui-ai — the abliterated (uncensored) Qwen3.8-27B backbone (Huihui-Qwen3.8-27B-abliterated).
• SargeDev — the distillation corpora: the question/state corpus (jev-distill-corpus-v3 and the per-domain jev-corpus-* split), and — for our solar LoRA — the typed-decision labels (solar-decisions-corpus-v4, 816,741 calibrated-distribution targets labeled by upstage/solar-decide-flash: a full relabel of v3 plus 120k new rows). The r32 half trains on the v3-lineage targets; the solar half trains on the v4 targets.
• Qwen — the base architecture.

Built on 2× NVIDIA GB10 (DGX Spark). Apache-2.0.