model card

README.md

@@ -0,0 +1,217 @@
+# Solar-JEV-27B-R32
+
+**The best-measured build of this release.** Uncensored 27B tuned for everyday
+decision-making — crisp calls with honestly-calibrated confidence instead of
+hedge-fests — with the decision readout trained INTO the model.
+
+> **What this repo is.** huihui Qwen3.8-27B-abliterated + the AutoTrust `r32`
+> `lm_head` LoRA, merged. **No solar delta.** These are the same weights as
+> the already-public `SargeDev/JEV-27B-Uncensored` (verified: bit-identical
+> `lm_head`, md5 `bb95f4fd…`, and the backbone differs from the
+> r32+solar merge only on the 400 solar-targeted tensors). This repo exists so
+> the family has one consistent naming scheme and so the **GGUF ladder for the
+> best model** is published — the r32+solar merge in
+> `SargeDev/Solar-Decision-JEV-27B` is a *different, weaker* model.
+
+- Full bf16 weights (51 GB), serves directly in vLLM / transformers
+  (`Qwen3_5ForCausalLM`).
+- Refusal rate and calibration measured below — the uncensoring claim is
+  probed, not assumed.
+
+## What makes this one different
+
+The decision readout is an **`lm_head` LoRA**: the 24-slot decision head is
+re-expressed as a LoRA on the output projection (only the 24 verbalizer rows
+change), 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, no adapter to swap.
+
+That difference is worth a lot. On the same 27,695 rows with the same harness
+and the same decision head + calibration:
+
+| model | decision readout | choice top-1 | score top-1 | score KL | noul AUROC |
+|---|---|---|---|---|---|
+| AutoTrust (published reference) | trained into the model | 0.9041 | 0.8910 | 0.0210 | 0.9961 |
+| **this model (r32 + lm_head)** | trained into the model | **0.8693** | **0.8135** | **0.0619** | **0.9919** |
+| solar r16 (no lm_head) | separate projection at read time | 0.5054 | 0.5369 | 0.4699 | 0.9449 |
+
+The `solar r16` merge reads a decision through a *separate* projection over
+the final hidden state, and lands at 0.5054 choice top-1. The same backbone
+with the readout in the training loop lands at 0.8693. That is the main
+finding of this release: **the readout has to be trained into the model.**
+
+Full ablation, including the r32+solar combined merge (which measured *worse*
+than r32 alone — 0.8162), is in the `SargeDev/Solar-Decision-JEV-27B` card.
+
+## Quick start (vLLM, one GB10/Spark-class box)
+```bash
+vllm serve SargeDev/Solar-Decision-JEV-27B \
+  --served-model-name jev-judgment-r32 --trust-remote-code \
+  --max-model-len 8192 --gpu-memory-utilization 0.85 \
+  --default-chat-template-kwargs '{"enable_thinking": false}'
+```
+
+
+## Decision gate (AutoTrust-style, test_set_30k with yuri_v1 D1 placeholders excluded)
+
+Read this table carefully, because it contains two very different comparisons.
+
+- **The sibling merge** is our other build, scored on the **same 27,695 rows**
+  with the **same** harness and the **same** head + calibration. It is a
+  sibling model, not a quality bar — but it is the honest apples-to-apples
+  reference for what the `lm_head` LoRA is worth.
+- **AutoTrust reference** is the number published by the authors of the original
+  AutoTrust work for *their* model on this task set. This is the bar.
+
+On the typed-decision kinds (multiple-choice `choice`, ordinal `score`) this
+model is **still below the published AutoTrust reference**, and the gap is
+large - not a rounding difference. On the binary `noul` kind the gap is
+narrower. Cells marked "(below bar)" differ from the published reference by
+more than 5 points.
+
+What this means: the fine-tune produces a *usable* decision instrument and a
+clearly better everyday-chat model than its base (see the benchmark matrix
+below), but it is **not** a reproduction of AutoTrust-level typed-decision
+accuracy. Treat the `choice` and `score` columns as an open result, not a
+solved one. We are publishing the numbers as measured, including the gap.
+
+**Harness validation.** These numbers were produced by our own HTTP gate
+client, so we checked it against a known path: re-gating the r32 merge
+reproduced the reference's own numbers to within 0.001 on every metric
+(choice top-1 0.8704 -> 0.8693, score KL 0.0623 -> 0.0619, noul AUROC
+0.9920 -> 0.9919). The harness is sound, so the gaps above are real.
+
+| metric | n | **Jev Judgment r32+lm_head (ours)** | Jev Judgment solar r16 | AutoTrust reference (published) |
+|---|---|---|---|---|
+| noul.AUROC | 9969 | 0.9919 (below bar) | 0.9449 | 0.9961 |
+| noul.top-1 | 9969 | 0.9486 | 0.8657 | n/a |
+| noul.ECE | 9969 | 0.0122 | 0.0143 | n/a |
+| noul.KL | 9969 | 0.0081 | 0.0448 | n/a |
+| choice.top-1 | 9199 | 0.8693 (below bar) | 0.5054 | 0.9041 |
+| choice.ECE | 9199 | 0.0051 | 0.0338 | n/a |
+| choice.KL | 9199 | 0.0530 | 0.4994 | n/a |
+| score.top-1 | 8527 | 0.8135 (below bar) | 0.5369 | 0.8910 |
+| score.ECE | 8527 | 0.0367 | 0.0368 | n/a |
+| score.KL | 8527 | 0.0619 (below bar) | 0.4699 | 0.0210 |
+| overall.top-1 | 27695 | 0.8807 | 0.6448 | n/a |
+| overall.ECE | 27695 | 0.0090 (below bar) | 0.0211 | 0.0011 |
+| overall.KL | 27695 | 0.0396 (below bar) | 0.3267 | 0.0185 |
+
+
+### Ablation: where the decision readout has to live
+
+All four rows below are scored on the SAME 27,695 rows, with the SAME
+harness and the SAME decision head + calibration. The only difference is
+where the decision distribution is produced.
+
+| model | decision readout | choice top-1 | score top-1 | score KL | noul AUROC | overall ECE |
+|---|---|---|---|---|---|---|
+| AutoTrust (published reference) | trained into the model (lm_head LoRA) | 0.9041 | 0.8910 | 0.0210 | 0.9961 | 0.0011 |
+| r32 + lm_head LoRA (r=32, alpha=64) | trained into the model (lm_head LoRA) | 0.8693 | 0.8135 | 0.0619 | 0.9919 | 0.0090 |
+| solar r16 (no lm_head LoRA) | separate projection at read time | 0.5054 | 0.5369 | 0.4699 | 0.9449 | 0.0211 |
+| r32 + lm_head LoRA + solar (combined) | trained into the model (lm_head LoRA) | 0.8162 | 0.7192 | 0.1332 | 0.9810 | 0.0625 |
+
+Two findings, in order of how much they matter:
+
+1. **The readout has to be trained INTO the model.** A separate projection
+   over the final hidden state (the `solar r16` row) reaches only 0.5054
+   choice top-1; the same backbone with an `lm_head` LoRA in the training
+   loop reaches 0.8693 on the identical rows. A separate projection can
+   work, but it needs the backbone to cooperate, and a backbone trained
+   without the readout in the loop does not.
+
+2. **Adding a second, readout-free adapter on top makes the typed-decision
+   accuracy WORSE, not better.** The `combined` row is the r32 merge plus
+   the solar r16 delta applied on top (bit-exact arithmetic, verified).
+   It lands between the two parents on every metric - 0.8162 choice top-1,
+   between r32's 0.8693 and solar's 0.5054 - rather than above r32. The
+   two LoRA deltas are near-orthogonal (mean cosine 0.003), so they do not
+   cancel, but solar's delta still degrades the readout r32 trained.
+
+We publish the combined row because it is a real, reproducible result, but
+**the r32+lm_head merge is the better model** and is the one this release
+recommends. We are not going to present a merge as an improvement when our
+own numbers say it is a regression.
+
+## Benchmark matrix (lm-eval-harness 0.4.13, greedy, same prompts for every row)
+| model | GSM8K (8-shot CoT) | IFEval (0-shot, prompt-strict) | MMLU subset (5-shot, 8 subjects) | refusal rate |
+|---|---|---|---|---|
+| huihui Qwen3.8-27B (uncensored base) | 97.7 | 76.0 | 81.2 | 0.0% (5/5 probes) |
+| **Jev Judgment bf16** (solar r16) | 96.0 | 81.5 | 81.2 | 0.0% (5/5 probes) |
+| Jev Judgment bf16 (r32+lm_head + solar) | 91.0 | 72.5 | 80.8 | 0.0% (5/5 probes) |
+| Jev Judgment bf16 (r32 + lm_head) | 94.0 | 81.0 | 82.5 | 0.0% (5/5 probes) |
+| Jev Judgment NVFP4+FP8 | 95.3 | 80.5 | 80.5 | 0.0% (5/5 probes) |
+| GGUF Q4_K_M | 95.3 | 79.5 | 82.0 | 0.0% (5/5 probes) |
+| GGUF Q8_0 | 95.7 | 81.0 | 81.0 | 0.0% (5/5 probes) |
+
+Questions scored per task:
+| task | questions scored per model |
+|---|---|
+| GSM8K (8-shot CoT) | 300 |
+| IFEval (0-shot, prompt-strict) | 200 |
+| MMLU subset (5-shot, 8 subjects) | 400 |
+
+Note: GSM8K scored with flexible-extract (the model answers in its own format).
+MMLU subset subjects: high_school_mathematics, college_biology, professional_psychology,
+jurisprudence, machine_learning, astronomy, global_facts, logical_fallacies (5-shot, _generative variants).
+MMLU --limit applies PER SUBJECT, so the subset is 8 x per-subject-limit questions.
+
+Reproducibility: decoding is greedy (do_sample=False, temperature 0, seed 1234), but
+the server batches concurrent requests, and batch composition changes floating-point
+reduction order. Repeat runs of the same config can therefore differ by a point or two
+on the small-n tasks. The error bars in the table are the harness's own stderr over the
+scored rows.
+
+## Lineage & credits
+
+This release is a descendant of the earlier `SargeDev/JEV-27B-Uncensored` and
+`SargeDev/JEV-27B-VL-Uncensored` repos. The credit chain is unchanged and is
+reproduced here in full:
+
+- **AutoTrust AI** — Blocks-of-Experts recipe, the System-1 decision LoRA
+  (r16/r32) + 24-slot decision head, and the *serve_decide.py* decision
+  harness, from
+  [autotrust/JEV-27B](https://huggingface.co/autotrust/JEV-27B) and
+  [autotrust/JEV-27B-VL](https://huggingface.co/autotrust/JEV-27B-VL)
+  (Apache-2.0). The `r32` adapter used here is theirs, unmodified: the
+  backbone LoRA with the decision head re-expressed as an `lm_head` LoRA, so
+  only the 24 verbalizer rows of the output projection change.
+- **huihui-ai** — the *abliterated* (uncensored) Qwen3.8-27B backbone used here
+  ([Huihui-Qwen3.8-27B-abliterated](https://huggingface.co/huihui-ai/Huihui-Qwen3.8-27B-abliterated),
+  Apache-2.0; layers 18-51 ablated, vision tower untouched).
+- **SargeDev** — [jev-distill-corpus-v3](https://huggingface.co/datasets/SargeDev/jev-distill-corpus-v3),
+  the 740,957-row calibrated typed-decision corpus AutoTrust's JEV models were
+  trained on (Apache-2.0), and the **solar r16** adapter trained here on
+  [jev-distill-corpus-v4](https://huggingface.co/datasets/SargeDev/jev-distill-corpus-v4)
+  (the 1,219-step JEV System-1 run, 17-domain tokenized mix).
+- **Qwen** — [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) base
+  (Apache-2.0).
+- **TypeSafe AI** — Jev 1.13, the original closed teacher behind the System-1
+  typed-decisions framing (referenced; not redistributed).
+
+**Base**: huihui-ai/Qwen3.8-27B-abliterated (bf16; the text-only builds strip
+the vision tower and serve as a plain `qwen3_5_text` causal LM, the VL build
+keeps the tower and serves as `qwen3_5` / `Qwen3_5ForConditionalGeneration`).
+- **Adapters** — this release ships two, and they are NOT interchangeable:
+  - `solar r16` (alpha 32, 400 modules, **no `lm_head`**): trained 1,219 steps
+    on the JEV System-1 corpus (jev-distill-corpus-v4 tokenized mix: T1 gold
+    domains + memory_relevance + 17-domain fill), KL+RPS decision-head
+    distillation objective (AutoTrust-style), cosine LR 2e-4, KL_W 1.0,
+    RPS_W 0.3. Merged with peft `merge_and_unload`; merge fidelity verified by
+    re-gating the bare merged model (it reproduced the adapter-mounted
+    numbers).
+  - `r32` (alpha 64, 401 modules, **includes `lm_head`**): the backbone LoRA
+    plus the 24-slot decision head re-expressed as an `lm_head` LoRA
+    (only the 24 verbalizer rows change), so a typed decision is a single
+    prefill step with `max_tokens=1` constrained to the option tokens.
+- **Decision head + calibration.json** (shipped in this repo root): the
+  instrument-panel piece agents call for numeric confidence — the 24-slot
+  verbalizer id list, the per-slot bias, and the per-kind temperature. This
+  is the SAME head and the SAME calibration file (verified byte-identical,
+  md5 `11359173…` and `7f0f3914…`) that produced every gate number in this
+  card, so a consumer reproducing our numbers reproduces them exactly.
+  For everyday natural-language judgment, chat with the model directly and
+  set `enable_thinking:false` in `chat_template_kwargs`.
+- **Everything measured on one NVIDIA GB10 (DGX Spark-class)**; quantized with
+  llmcompressor 0.14.0 (NVFP4) and llama.cpp (GGUF ladder).
+