전체 검색

AI 모델 2119

• roberta-mixed-detector↗

  andreas122001

  This project is related to a bachelor's thesis with the title "*Turning Poachers into Gamekeepers: Detecting Machine-Generated Text in Academia using Large Language Models*" (see here) written by *Nicolai Thorer Sivesind* and *Andreas Bentzen Winje* at the *Department of Computer Science* at the *Norwegian University of Science and Technology*. It contains text classification models trained to distinguish human-written text from text generated by language models like ChatGPT and GPT-3. The best models were able to achieve an accuracy of 100% on real and *GPT-3*-generated wikipedia articles (4500 samples), and an accuracy of 98.4% on real and *ChatGPT*-generated research abstracts (3000 samples). **NOTE**: the hosted inference on this site only works for the RoBERTa-models, and not for the Bloomz-models. The Bloomz-models otherwise can produce wrong predictions when not explicitly providing the attention mask from the tokenizer to the model for inference. To be sure, the pipeline-library seems to produce the most consistent results.

• EsoTest_GGUFs↗

  Esobold

  - **GhostShip-31b-Q4_K_M.gguf** - FM nearswap merge: Gemma4-Gutenberg-31B-Heretic 0.0007 linear across layers (avg 4e-4). Same family as Robinson/BenGunn, ~4x peak strength.

• Smaug-Llama-3-70B-Instruct-abliterated-v3↗

  failspy

  This is abacusai/Smaug-Llama-3-70B-Instruct with orthogonalized bfloat16 safetensor weights, generated with a refined methodology based on that which was described in the preview paper/blog post: 'Refusal in LLMs is mediated by a single direction' which I encourage you to read to understand more. TL;DR: This model has had certain weights manipulated to "inhibit" the model's ability to express refusal. It is not in anyway _guaranteed_ that it won't refuse you, understand your request, it may still lecture you about ethics/safety, etc. It is tuned in all other respects the same as the original 70B instruct model was, just with the strongest refusal directions orthogonalized out. **TL;TL;DR;DR: It's uncensored in the purest form I can manage -- no new or changed behaviour in any other respect from the original model.**

• LFM2.5-cmf↗

  infosave

  LiquidAI's LFM2.5 is a hybrid: most layers mix with a short gated convolution and only a few carry full attention, so the state a token needs is small and constant where the convolution runs. These are those checkpoints in the CMF container — a single memory-mapped file read by `cortiq`, a Rust binary with no ML framework under it. GPU via Vulkan/Metal/DX12 with a CPU fallback; NVIDIA, AMD, Intel and Apple silicon read the same file. All three are 4-bit tiled with ladder scales (`q4tp`), quantized straight from the bf16 checkpoints. Steady-state decode, `cortiq bench --core`, single stream, cortiq **0.5.99+** (the whole-token graph learned this family's short-conv mixer and its sigmoid-routed MoE in 0.5.98/0.5.99 — earlier versions decode it an order of magnitude slower).

• Yi-1.5-9B-Chat-16K-GGUF↗

  QuantFactory

  Yi-1.5 is an upgraded version of Yi. It is continuously pre-trained on Yi with a high-quality corpus of 500B tokens and fine-tuned on 3M diverse fine-tuning samples. Compared with Yi, Yi-1.5 delivers stronger performance in coding, math, reasoning, and instruction-following capability, while still maintaining excellent capabilities in language understanding, commonsense reasoning, and reading comprehension. Model | Context Length | Pre-trained Tokens | :------------: | :------------: | :------------: | | Yi-1.5 | 4K, 16K, 32K | 3.6T

• plant-disease-dinov2-small↗

  rodynaemad

  Identifies **38 crop–condition classes** (26 diseases and 12 healthy classes across 14 crops) from a photo of a single leaf. Built as the open disease-scanning model for BloomTech, my AI + IoT smart-agriculture graduation project. PlantVillage contains several photos of the **same physical leaf**. If photos are split randomly, near-duplicates of a test leaf end up in training, which inflates test accuracy. The dataset authors' `leaf-map.json` links 40,490 photos to 7,524 physical leaves. I split the data so that **every photo of a given leaf lands on only one side** (stratified, leaf-grouped 80/20 split), and asserted that no leaf crosses the split. For this model the inflation from a naive split is small (99.1% vs 98.9%), but the leaf-level number is the one to trust. The remaining 13,815 photos have no leaf mapping and are split photo by photo. `onnx/model_with_features.onnx` (88 MB, fp32) and `onnx/model_with_features_quantized.onnx` (24 MB, int8) take `pixel_values` (1×3×224×224, ImageNet-normalized) and return `logits` and the 768-d `features` used for the out-of-distribution check. The fp32 export matches PyTorch exactly. The int8 copy, which powers the Space, scores **98.7% accuracy (98.1% macro F1)** on the full leaf-level test set. Example with onnxruntime-web in the Space's `index.html`.

• bert-base-arabertv02-twitter↗

  aubmindlab

  AraBERTv0.2-Twitter-base/large are two new models for Arabic dialects and tweets, trained by continuing the pre-training using the MLM task on ~60M Arabic tweets (filtered from a collection on 100M). The two new models have had emojies added to their vocabulary in addition to common words that weren't at first present. The pre-training was done with a max sentence length of 64 only for 1 epoch. **AraBERT** is an Arabic pretrained language model based on Google's BERT architechture. AraBERT uses the same BERT-Base config. More details are available in the AraBERT Paper and in the AraBERT Meetup

• pocket-tts-GGUF↗

  cstr

  Pocket TTS is a lightweight (~100M param) continuous-latent autoregressive TTS model from Kyutai, based on the CALM paper (arXiv:2509.06926). Unlike codebook-based TTS models, Pocket TTS emits continuous float vectors — no discrete tokens, no softmax sampling: - **FlowLM backbone** — causal transformer (1024D, 16 heads, 6 layers; 24 layers for the French preview, RoPE, GELU) operating at 12.5 Hz - **Consistency head** — SimpleMLPAdaLN (512D, 6 ResBlocks) with timestep embedding → one-step LSD decode → 32-dim continuous latent vectors - **Mimi VAE decoder** — SEANet upsample convolutions + 2-layer transformer → 24 kHz PCM - **Mimi VAE encoder** *(voice-cloning builds only)* — SEANet downsample + 2-layer transformer + speaker projection → reference conditioning - **Text tokenizer** — SentencePiece BPE (4000 vocab, embedded in GGUF) Single GGUF file — no separate codec companion needed (Mimi weights and the tokenizer are embedded). Pocket TTS is zero-shot: it clones the timbre of a short reference clip. The **voice-cloning builds** (the files *without* `novc` in the name) embed the Mimi VAE **encoder** and speaker projection needed to condition on a reference; the English-only `novc` builds omit that encoder and are ~20 MB smaller.

• GLM-5.3-Flash-ewaste-edition-GGUF↗

  SixVolts

  This is a quantization of GLM-5.3-flash that is optimized to work better on older hardware. This quant was built using i-matrix GGUF quantizations with the same principle as the GLM-5.2 and GLM-5.3 e-waste editions: **K-quants for the routed experts, no codebook i-quants**, because on pre-AVX-512 CPUs and previous-generation datacenter GPUs (MI100 / gfx908) expert dequantization, not bandwidth, is the decode bottleneck. **Format: mainline llama.cpp `glm5-next`** (merged 2026-09-30, ggml-org/llama.cpp#27773). These files do **not** load in builds that only know the earlier `glm5next` arch from the unmerged PR #27754 (unsloth's September files are that format). Any llama.cpp from 2026-10 on loads them. The NextN/MTP head (blk.45) is in every file for when mainline's MTP support (#27917) lands. All are imatrix-quantized with unsloth's GLM-5.3-Flash importance matrix (88 chunks × 9216 tokens) from an **F32 intermediate** of the FP8 checkpoint (`fp8 × block scale` in F32; a BF16 intermediate re-rounds most weights because the scales are arbitrary floats). Tensors that mainline's quantizer keeps unquantized for this arch (hyper-connection mixers, DSA indexer, KDA gates/betas, `attn_kv_a_mqa`) are F32 in every file; `attn_q_a/q_b` and `nextn.eh_proj` are Q8_0 (mainline's floor). The MTP head's experts are Q4_K in every variant.

• cobalt-seeded-rl-base-ramp25-stoppen-gen4k-ep2-ncp5-base-q3c↗

  agurung

  Trained and validated on the **cobalt-train ≤2/64 frontier** (canonical clean_eval prompts): 1833 train / 112 held-out val problems the base model solved on at most 2 of 64 samples under the iid_canonical@64 hardness scan. Val evals sample at temperature 1.0 (matching the clean_eval frontier eval). Reward signal: **binary code-correctness** (1.0 if the generated program passes the problem's tests, otherwise 0.0). This checkpoint is the **`main`** revision (git branch) of the repo, with the model at the repo root — load it directly, no subfolder: