kortexa-ai/shingi-bonsai-2-27b-v0.2
Shingi — 審議 — Bonsai 2 27B decision adapter v0.2
Shingi is an experimental local decision model for request-defined choices, probabilities, and ordinal scores. It reads all candidate logits from Bonsai and returns structured results without generating an explanation.
This repository contains a 132.03 MiB FP32 LoRA adapter, calibrated readout parameters, evaluation charts, and a checksum manifest. Download the unchanged ternary Bonsai base separately. Use the Shingi CUDA runtime with the pinned Prism fork; this adapter is not a standalone model or a generic Transformers/PEFT adapter. Created using Bonsai by Prism ML.
What changed
Version 0.2 trains a fresh adapter using six sources with CC BY, MIT or CC0 terms. ARC and BoolQ are excluded from training, development selection and calibration. No previous adapter weights, optimizer state or fitted calibration were reused. The new adapter is Apache-2.0; the historical v0.1 adapter keeps its CC BY-SA 4.0 license.
Rank-8 LoRA with alpha 16 adds 34,603,008 trainable parameters to the gate, up and down MLP projections of all 64 layers. Original ternary weights, scales and transforms remain unchanged. The selected checkpoint is update 559, with separate probability calibration. No adapter merge or additional release quantization was performed.
Choice-map reordering is stable because keys are sorted before inference. Renaming choices, changing their wording, or varying the underlying prompt order can still change the result. Score levels retain their supplied order.
Run on CUDA
Tested targets are Linux x86-64, RTX 4090, and RTX PRO 6000 Blackwell. Install Python 3.11+, uv, Git, CMake, a C++17 compiler, the Hugging Face CLI, and CUDA 12.8+ with a compatible driver. The measured build used CUDA 13.0.88.
git clone --branch v0.2.0 https://github.com/kortexa-ai/shingi.git
cd shingi
uv sync --locked
export PATH=/usr/local/cuda/bin:$PATH
bash scripts/setup_cuda.sh
hf download prism-ml/Ternary-Bonsai-2-27B-gguf \
Ternary-Bonsai-2-27B-PQ2_0.gguf LICENSE NOTICE.txt \
--revision 6ed5e12bf84b7a63069882c91dd9e9218647d17b \
--local-dir artifacts/base
hf download kortexa-ai/shingi-bonsai-2-27b-v0.2 \
--revision v0.2.0 \
--local-dir artifacts/shingi-v0.2
nvidia-smi --query-gpu=name,uuid,memory.free --format=csv
export CUDA_VISIBLE_DEVICES=GPU-YOUR-FULL-UUID
uv run --locked shingi \
--model artifacts/base/Ternary-Bonsai-2-27B-PQ2_0.gguf \
--adapter artifacts/shingi-v0.2/adapter.gguf \
--calibration artifacts/shingi-v0.2/calibration.jsonUse a complete GPU UUID from nvidia-smi. Startup verifies model and adapter checksums against calibration. The API binds to 127.0.0.1:8765.
curl -s http://127.0.0.1:8765/v1/systemone \
-H 'Content-Type: application/json' \
-d '{"model":"shingi","state":"The parcel arrived damaged. Please replace it.","questions":{"route":{"type":"choice","instructions":"Select the handling team.","criteria":{"returns":"Damage and replacements","billing":"Charges and refunds","shipping":"Delivery tracking"}}}}'The API contract describes Choice, Noul (truth probability), Score, and real TypeSafe SDK support. Shingi needs no Kortexa service or private runtime. It serves sequentially, reads every candidate logit, and rejects oversized inputs without truncation. The context limit is 16,384 tokens. Up to 52 choices need one pass; 53–255 use approximate chunk-and-anchor scoring with multiple passes.
Training and data
The selected checkpoint saw 4,466 prompts from 2,233 distinct records and 1,073,860 input tokens. The prepared pool held 2,284 records from Banking77, CLINC150, MMLU, HelpSteer2 helpfulness, Measuring Hate Speech and Civil Comments. JevBench supplied transformations at revision 002ad22de8db2df5e0eb898b3da8072dbd4af4de. The pool reused 378 permitted earlier training records. All published validation/test states and prior project evaluation records were excluded from training. New development and calibration data exclude every earlier project split by ID and canonical state hash.
Two deterministic option permutations were prepared per training record; 102 prompts over 1,024 tokens were excluded without truncation. Training used candidate cross entropy against human vote distributions where available, otherwise the gold label: batch one, accumulation eight, AdamW, learning rate 5e-5, 32 warmup updates, cosine decay and gradient clipping at 1. The run completed one prepared epoch, 559 updates, with zero skipped optimizer steps. No hosted Jev predictions were training targets.
Checkpoint selection used minimum uncalibrated NLL on 192 development records. Update 559 reduced development NLL to 0.626760. Its accuracy was 75.52%, below update 512's 76.56%; the selection rule used NLL and was set before training. Native development predictions agreed with the selected differentiable model on all 192 selected answers, with mean probability TVD 0.001770. This is measured agreement, not exact numerical equivalence. The training record and bundled training-provenance.json retain source audit and measurements.
The canonical Choice readout was fixed before the run. Calibration uses a separate 180 records: 150 Choice/Score and 30 Noul. The new adapter's global temperature is 1.5; Noul additionally divides by 1/3, with zero log-odds bias. The unchanged base has its own calibration. The old adapter retains its original 240-record calibration for the matched comparison.
Weights and calibration were frozen before drawing the fresh 1,500-record test, 100 per source. It excludes 10,708 previous IDs and 10,708 state hashes. Nine sources were excluded from every v0.2 fitting stage: ARC Challenge, BoolQ, LEDGAR, GoEmotions, MNLI, SST-5, FEVER Evidence, SMS Spam and ChaosNLI. Their absence from adaptation does not prove absence from upstream pretraining; that overlap is unknown. No test result selected or tuned the release.
Accuracy and probability quality
On the fresh test, Shingi v0.2 scores 1124/1,500 (74.93%) versus 1002/1,500 (66.80%) for unchanged Bonsai on the RTX PRO 6000. The paired gain is +8.13 percentage points, with a descriptive 95% bootstrap interval of +6.00 to +10.27 points.
The new adapter is not better on every metric. ECE worsens on both cards. On the 6000, HelpSteer2 falls from 48 to 45 correct answers, SMS Spam from 99 to 96, and ChaosNLI from 60 to 57, each out of 100. The 4090 has the same HelpSteer2 and SMS losses; ChaosNLI falls from 60 to 58. NLL, Brier and expected-score error improve overall. The accuracy intervals against v0.1 include zero, so these small aggregate gains do not establish a clear advantage.
Each adapter retains its own frozen calibration; this compares the packaged models rather than isolating weight changes. No test-time adjustment was made.
On the nine sources excluded from v0.2 fitting, accuracy improves from 615/900 (68.33%) to 670/900 (74.44%) against the base (paired interval +3.67 to +8.56 points). This does not establish general unseen-domain reliability. GoEmotions remains weak at 27/100.
The full report retains every measured regression, denominators, intervals and per-source results. The machine-readable evidence also includes uncalibrated old/new comparisons. Score accuracy uses the modal level; expected-score error and distance to human vote targets are reported separately.
External benchmarks
The same frozen weights, calibration and prompt answered all 8,122 questions from three additional evaluation-only suites on the RTX PRO 6000. These results did not select or tune the release.
Spatial reasoning remains weak: This/That's only_way family scores 23.37%, plan_survives 30.60%, and distance_band 32.20%. Across its 500 stochastic questions, analytic expected accuracy is 51.01%, versus a Bayes ceiling of 85.88%; the full-set uniform-random baseline is 34.76%.
The Jev-Omni model card reports 87.57% macro on DecisionBench Medium and 86.15% macro on public JevBench. Those are third-party reports, not our measurements. Its complete training corpus, frozen evaluation manifest, predictions and exact scoring code were not available in that release. The comparison does not establish protocol equivalence or statistical significance. No hosted Jev model was benchmarked. Public JevBench contains 231 of 534 full-suite tasks; this is not an official full-suite composite score.
Choice and Score accuracy use argmax; Noul uses P(yes) ≥ 0.5. Rounding Score's expected level instead gives DecisionBench macro 80.50% Medium and 59.55% Hard; Jev-Omni's exact Score rule is unknown. Thirteen Choice questions per subset exceed 52 options and use the existing approximate readout, up to 120 options. Canonical map reversal has zero flips in 300 pairs; changing raw prompt order flips 28/100 This/That, 18/100 DecisionBench and 9/100 JevBench answers.
The external report, metrics, provenance and validation retain probability quality, latency, memory, source revisions, overlap checks and complete audit identities. No exact or normalized state overlap was found against 11 earlier Shingi data files; semantic overlap and upstream pretraining exposure remain unknown.
CUDA latency and memory
The mixed workload uses 300 fixed records, 20 per source, repeated three times after ten warmups. Synthetic curves use ten measured repeats per case. Short HTTP timings use 30 localhost SDK requests after three warmups. Request wall time includes tokenization, GPU memory checks and all candidate passes. Loading is separate; model initialization excludes artifact checksum hashing.
The RTX PRO 6000 retained six running production services during inference; the RTX 4090 retained one unrelated GPU process. Old and new adapters used the same source, native executable and workload, with sequential runs. Other workload activity and filesystem cache were uncontrolled. Small timing changes are not evidence of a speedup. Hardware was an Intel Core i9-14900K, CUDA 13.0.88 and driver 610.43.02, with power limits of 450 W (6000) and 480 W (4090).
Memory is the increase in device allocation over the pre-load baseline, sampled every 100 ms with a 16,384-token context and Q8 KV cache. A brief peak can be missed, and the measurements are not isolated to one process. Host RSS was sampled only after readiness, so it does not establish a host load-time peak. CUDA memory figures do not establish feasibility on a 24 GB Mac.
Cross-GPU agreement
The RTX 4090 scores 1,125/1,500 (75.00%) versus 1,004/1,500 (66.93%) for unchanged Bonsai. Shingi agrees across GPUs on 1496/1,500 decisions (99.73%), with mean probability TVD 0.00280. The base agrees on 1493/1,500 (99.53%), with mean TVD 0.00485. Both pass the predeclared 99% agreement and 0.01 mean-TVD thresholds.
This is not exact numerical equivalence. Three requests per model select a different intermediate chunk winner and therefore a different final comparison prompt. All static prompts and candidate token IDs match, and every adaptive path was checked by replay. The new adapter's maximum individual TVD is 0.34677, on a 77-choice Banking77 request. Small hardware-dependent logit changes can produce large probability shifts in this approximate method.
Both cards pass 12/12 synthetic context probes and all live SDK checks, including 52, 53 and 255 choices, map reordering and explicit oversized-input errors. Raw input-order changes flip 14/200 new-adapter answers on each card. Canonical map reordering has zero flips on both cards; this is an interface property, not learned model invariance.
Limitations
This is one small adaptation experiment, primarily on English benchmark text. It does not establish general reliability, fairness, or robustness to prompt injection. Evaluate the model on your own domain before delegating decisions. Confidence fields describe distribution concentration, not a guarantee of correctness. Calibration can change under a new workload.
The original release test has only 100 records per source and gives every source equal weight. Order sorting does not remove label or wording sensitivity. The synthetic context tests check simple planted-fact retrieval, not real-document comprehension. More than 52 choices use an approximate probability composition. Images, audio and video are unsupported. No M3/M4 Mac performance, concurrency throughput, hosted Jev comparison, or generation-token speed is claimed.
Files, identity and license
adapter.gguf: SHA-2566f4c2b7434b5ef6e0bb83b6f9c196de00c850a79b9380175e06a1df2ae1b7dd2.- Required base:
Ternary-Bonsai-2-27B-PQ2_0.gguf, revision6ed5e12bf84b7a63069882c91dd9e9218647d17b, SHA-2563907dc1658db1f78a9826bf8d5bcb8dc65db0d466388937af57f2294fae62ec1. - Native runtime: PrismML-Eng/llama.cpp revision
d8f26eec76da6d09bb708bcba51ef64b8cd868a3. calibration.jsonbinds calibration to both weight hashes and the readout version.protocol.jsonrecords the pre-test freeze;manifest.jsonlists every bundled file and checksum.evaluation/contains aggregate metrics and data identifiers;figures/contains PNG and SVG charts. Raw datasets and prompts are not redistributed.
The v0.2 adapter, Shingi-authored documentation and software are Apache-2.0. The unchanged Bonsai base is Apache-2.0 and the Prism runtime is MIT. Third-party materials retain their own terms. The six fitting sources use CC BY 4.0, CC BY 3.0, MIT or CC0; their attribution and pinned license documents are retained. Evaluation-only datasets have separate terms. The historical v0.1 weights remain CC BY-SA 4.0. See the source review, NOTICE and bundled licenses/.
Bonsai is by Prism ML, based on Qwen by Alibaba Cloud. JevBench transformations are by Praveenrajus / uspraveen and retain the underlying sources' attribution. OpenJev's Apache-licensed helper and the MIT TypeSafe adapter informed the readout and interface. Shingi includes no OpenJev weights and is not endorsed by those upstream projects.