Open-Weight Models

Native MLX conversion of ewin-reg/WeMM-Embedding-2B-Quantized — the hybrid FP8-protected attention + group-16 INT4 checkpoint — that preserves its mixed-precision allocation instead of flattening everything to one precision. The source checkpoint is curvature-guided: it deliberately keeps the precision-sensitive path at ~8-bit (FP8 E4M3) and compresses only the quantization-robust masses to INT4. Uniform 4-bit quantization destroys that — the source README itself documents softmax attention collapse (DKL >= 7.50) when Q/K go to 4 bits. This build mirrors the Two deviations from the source, both improvements or forced by MLX: - 8-bit group-64 is finer than the source's per-tensor-scalar FP8…

A bilingual vision-language model built on Keural Mid, a 183.6M-parameter vision encoder trained from random initialization — not a CLIP, ViT or DINO derivative. This repository contains the trained bridge only: a projector and LoRA adapters, 37 MB total. The encoder and decoder load from their own repositories. A vision-language model can post respectable scores while barely using the image — "what colour is the banana?" is answerable from language priors alone. Every benchmark below was therefore run twice: once normally, once with the visual tokens removed and everything else held constant. The delta is the encoder's contribution. It is the one number a language model cannot manufacture.…

A 183.5M-parameter image encoder built from scratch — adaptive tokenization, hierarchical concept tokens, and continuous 2D positional encoding. Not a CLIP/ViT fine-tune. Measured at 384×384 with tokenbudget=512, matching training. Accuracy trails same-size peers — see Limitations and Evaluation for why, and for the higher-k retrieval results where the gap closes. The encoder is frozen and used as-is inside a bilingual (English + Korean) Only a projector (8.9M) and LoRA adapters (5.0M) are trained on top; the decoder is Qwen2.5-7B-Instruct. A vision-language model can score respectably while barely using the image, because many questions are answerable from language priors alone. Each…

Model · Text classification

kev-0.8b-code-verify-v1

James

A fine-tune of jaredpalmer/kev-0.8b (first fine-tune): a Jev-style decision model that judges whether code written by a cheap coding model is correct, so a router can keep the answer local or escalate it. One forward pass, a calibrated probability, no generated text. Trained on execution-labelled data: coder attempts at HumanEvalPack (Python) tasks, labelled by running each task's hidden test suite. The model never sees the hidden tests; its input is what a router can compute itself. The fine-tune binds these exact strings. One noul question (probability that the statement holds): State (an object; Kev renders it as key: value lines): generatededgecasetests comes from a small model writing…

Model · Text classification

kev-0.8b-code-verify-v2

James

A fine-tune of jaredpalmer/kev-0.8b (current best local judge: weak-coder data added): a Jev-style decision model that judges whether code written by a cheap coding model is correct, so a router can keep the answer local or escalate it. One forward pass, a calibrated probability, no generated text. Trained on execution-labelled data: coder attempts at HumanEvalPack (Python) tasks, labelled by running each task's hidden test suite. The model never sees the hidden tests; its input is what a router can compute itself. The fine-tune binds these exact strings. One noul question (probability that the statement holds): State (an object; Kev renders it as key: value lines): generatededgecasetests…

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019). - PEFT 0.21.0

Model

kev-4b-actions

Aaaaflolep

LoRA for Kev (Qwen3.5-4B-Base + rank-16 LoRA + pointer head, the /v1/systemone contract) that picks the next browser action in the custombrowser harness: from a store's homepage, find a named product and open its page. It replaces only the action question. The done check is a separate LoRA, Flolep2607/kev-4b-guard@v2. 14 tasks x 2 runs on obscura, starting at the store's homepage, sites never trained on. Corrected 2026-10-14: every check is exact now (the final address must end with the target product's path). Two timber products are not sold on their site (Bunnings, NZ Timber): those tasks pass only when the agent gives up on its own. The first version of this card used looser checks that…

Model · Text classification

kev-4b-code-verify-v1

James

A fine-tune of jaredpalmer/kev-4b (4B, trained on a bf16 backbone on a single L4): a Jev-style decision model that judges whether code written by a cheap coding model is correct, so a router can keep the answer local or escalate it. One forward pass, a calibrated probability, no generated text. Trained on execution-labelled data: coder attempts at HumanEvalPack (Python) tasks, labelled by running each task's hidden test suite. The model never sees the hidden tests; its input is what a router can compute itself. The fine-tune binds these exact strings. One noul question (probability that the statement holds): State (an object; Kev renders it as key: value lines): generatededgecasetests comes…

Why the disagreement is the feature, not the bug: kev downgrades high-frequency auto-recovering alerts (the #1 source of alert fatigue) and escalates orchestration-core outages that static rules treat as routine. - None-injection (p=0.1): teaches the head to abstain, not everything deserves a label We ran the obvious next experiment: fine-tune it on its own production data (5-fold grouped CV, 4-arm temperature calibration). Result: 0 gain over the deployed version — and we did not deploy it, per preset criteria defined before the experiment. That's the discipline this model is proud of: judge the upgrade on pre-registered criteria, not vibes. Training receipts and the arbitration table are…

Working draft v6 (2026-10-07), experimental technical report, not peer reviewed. The roughly twofold joint-KL ratio in §5.7 is a provisional result while a matched-denominator recheck finishes; other numbers may still change. kev-d3-paper-draft-v6.pdf (current) Per-module sub-3-bit sensitivity of two Qwen3.5 hybrid decision models (Gated DeltaNet linear attention interleaved 3:1 with full attention), measured as option-distribution KL under binary / ternary / 3-bit / 4-bit GPTQ, with a multiple-choice knapsack priced in the target runtime's storage format (HyBitQ) that turns the measurements into bit allocations. Related release: Jakevin/kev-4b-ternary-mlx (revision v2.0 uses this…

Fine-tuned from jaredpalmer/kev-4b to recommend math journals from a paper’s title and abstract. The model scores all 95 journal names in alphabetical order, without supporting papers or journal descriptions. This package contains the epoch-2.5 LoRA adapter, decision head, and tokenizer. It requires the pinned Qwen3.5-4B-Base backbone. Journal ranking does not require an embedding model or reference corpus. Both models received identical requests with all 95 journal names. Top 3 measures whether the paper's publication journal appears among the first three recommendations. The checkpoint was fixed using validation before this evaluation. Test papers were excluded from training and…

Model

kev-queue-v4-4b

Fzkun

基于 KEV 框架微调的中文排队位置决策模型。输入现有队列和明确的插入规则，输出候选位置的概率分布及 positionN 选项。模型不生成自由文本，也不用于医疗诊断或判断患者病情紧急程度。 本仓库包含 LoRA 适配器和 KEV 决策指针头，不包含完整的基础模型权重。运行需要指定版本的基础模型、KEV 框架、分词器和推理代码；直接用通用文本生成管线加载适配器，无法复现本模型的决策输出。 公开网页是 确定性规则演示，不调用微调模型。本地版本通过服务端代理调用内网模型 API，并对照规则答案。网页后来新增的普通报到后移策略不属于本模型原有训练范围，不能用网页中的规则计算结果代表模型准确率。 原有患者的相对顺序保持不变。候选名称 position16 已表示 第 16 位，不要再加 1；模型置信度不是位置编号，也不是正确率保证。 call-system 的普通报到流程还会按排队号计算候选位置，跳过候选落点后连续的已调整患者，在第一个未调整患者之前停止。该流程与回诊、献血、家优的搜索算法不同，不能统一追加到所有策略中。 公开网页已增加这个规则的模拟场景，但 本模型尚未针对普通报到、调整状态、跨班次或加号规则重新训练。这些操作应由业务代码处理。网页保留源码 index < len(list)-1 的后移扫描边界，仅演示同班次普通号。 此次实际训练采用 多任务监督微调（SFT）和对称的候选排列一致性 KL 约束，没有进行 RL 或 RLCD。 建议在 Docker 中使用指定版本的 KEV 源码及兼容依赖。已验证部署采用 PyTorch 2.8.0、transformers…

Each LoRA is its author's work, under its own licence; this folder holds the builds the Studio's two engines load. SOURCE.md in each folder names the author, the original file and what our build changed. Powered by MiniMax H3.

This model is a fine-tuned version of Kkonjeong/wav2vec2-base-korean on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 10 - mixedprecisiontraining: Native AMP - Transformers 5.18.0 - Pytorch 2.11.0+cu130 - Datasets 4.8.5 - Tokenizers 0.23.2

Model · Audio classification

keyword

Yim

This model is a fine-tuned version of Kkonjeong/wav2vec2-base-korean on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 4 - evalbatchsize: 4 - lrschedulertype: linear - numepochs: 10 - mixedprecisiontraining: Native AMP - Transformers 5.18.0 - Pytorch 2.11.0+cpu - Datasets 4.8.5 - Tokenizers 0.23.2

This model is a fine-tuned version of Kkonjeong/wav2vec2-base-korean on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 10 - mixedprecisiontraining: Native AMP - Transformers 5.18.0 - Pytorch 2.11.0+cu130 - Datasets 4.8.5 - Tokenizers 0.23.2

This model is a fine-tuned version of Kkonjeong/wav2vec2-base-korean on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 10 - mixedprecisiontraining: Native AMP - Transformers 5.18.0 - Pytorch 2.11.0+cpu - Datasets 4.8.5 - Tokenizers 0.23.2

This model is a fine-tuned version of Kkonjeong/wav2vec2-base-korean on the None dataset. It achieves the following results on the evaluation set: - evalloss: 0.9620 - evalaccuracy: 0.9474 - evalruntime: 18.9185 - evalsamplespersecond: 1.004 - evalstepspersecond: 0.159 The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 10 - mixedprecisiontraining: Native AMP - Transformers 5.18.0 - Pytorch 2.11.0+cpu - Datasets 4.8.5 - Tokenizers 0.23.2

This model is a fine-tuned version of Kkonjeong/wav2vec2-base-korean on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 10 - mixedprecisiontraining: Native AMP - Transformers 5.18.0 - Pytorch 2.11.0+cpu - Datasets 4.8.5 - Tokenizers 0.23.2

This model is a fine-tuned version of Kkonjeong/wav2vec2-base-korean on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 10 - mixedprecisiontraining: Native AMP - Transformers 5.18.0 - Pytorch 2.11.0+cu130 - Datasets 4.8.5 - Tokenizers 0.23.2

This model is a fine-tuned version of Kkonjeong/wav2vec2-base-korean on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 10 - mixedprecisiontraining: Native AMP - Transformers 5.18.0 - Pytorch 2.11.0+cpu - Datasets 4.8.5 - Tokenizers 0.23.2

This model is a fine-tuned version of Kkonjeong/wav2vec2-base-korean on the None dataset. It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 8 - evalbatchsize: 8 - lrschedulertype: linear - numepochs: 10 - mixedprecisiontraining: Native AMP - Transformers 5.18.0 - Pytorch 2.11.0+cpu - Datasets 4.8.5 - Tokenizers 0.23.2