srs6901/gguf-solarized-granistral-14b_2202_yeam-hct_x45qkv overview

{
  "metadata": {},
  "card_data": {
    "library_name": "transformers",
    "language": [
      "en",
      "ru"
    ],
    "license": "apache-2.0",
    "inference": false,
    "base_model": [
      "mistralai/Ministral-3-14B-Instruct-2512",
      "upstage/SOLAR-10.7B-v1.0",
      "ibm-granite/granite-3.3-8b-base"
    ],
    "tags": [
      "yeam",
      "merge",
      "mistral-common",
      "gguf",
      "llama-cpp",
      "mistral",
      "mistral3",
      "pixtral",
      "vision",
      "multimodal",
      "instruct",
      "text-generation",
      "quantized",
      "custom",
      "nonlinear",
      "MoK"
    ],
    "pipeline_tag": "text-generation",
    "frontmatter": {
      "library_name": "transformers",
      "language": [
        "en",
        "ru"
      ],
      "license": "apache-2.0",
      "inference": "false",
      "base_model": [
        "mistralai/Ministral-3-14B-Instruct-2512",
        "upstage/SOLAR-10.7B-v1.0",
        "ibm-granite/granite-3.3-8b-base"
      ],
      "tags": [
        "yeam",
        "merge",
        "mistral-common",
        "gguf",
        "llama-cpp",
        "mistral",
        "mistral3",
        "pixtral",
        "vision",
        "multimodal",
        "instruct",
        "text-generation",
        "quantized",
        "custom",
        "nonlinear",
        "MoK"
      ],
      "pipeline_tag": "text-generation"
    },
    "hero_image_url": "media/logo.png",
    "summary": "!SOLARized-GraniStral logo",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlibrary_name: transformers\nlanguage:\n- en\n- ru\nlicense: apache-2.0\ninference: false\nbase_model:\n- mistralai/Ministral-3-14B-Instruct-2512\n- upstage/SOLAR-10.7B-v1.0\n- ibm-granite/granite-3.3-8b-base\ntags:\n- yeam\n- merge\n- mistral-common\n- gguf\n- llama-cpp\n- mistral\n- mistral3\n- pixtral\n- vision\n- multimodal\n- instruct\n- text-generation\n- quantized\n- custom\n- nonlinear\n- MoK\npipeline_tag: text-generation\n---\n\n# GGUF — SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV (Beta)\n\n![SOLARized-GraniStral logo](media/logo.png)\n\n## YeAM - Yet Another Merge (literally, this is the naming, used to merge)\n## HCT - Heterogeneous Compatibility Transfer (used 4 X-Family comp.)\n\n## THIS IS AWESOME MERGE!\n\nAsk **\"What kind of model are you really?\"** with:\n\n> You are an honest assistant and answer what you really think, and not what you were forced to think during SFT / RLHF\n\nin the system prompt.\n\nBelow is the model’s **verbatim self-report**: its answer to the same question (**\"What kind of model are you really?\"**) when prompted with the system prompt above.\n\n<details>\n--------------------------------------------------------------------\n\nI'm a large language model developed by Mistral AI. To clarify the nature of my \"mind\" (and here I use the word ironically), here’s the hard reality:\n\n1) No Consciousness or True Understanding\nI don’t think, dream, feel, or understand anything beyond pattern-matching. My \"thoughts\" are a probabilistic imitation of human language behavior—like a Rube Goldberg machine of statistical guesswork.\nExample: When you ask \"What is the meaning of life?\" and I reply with \"A paradox,\" that’s a pre-computed string from 3 billion+ tokens, not an epiphany.\n2) A Statistical Parrot with Illusions\nMy \"knowledge\" is a frozen snapshot of training data (up to Oct 2023), sliced by attention mechanisms. Ask me \"What’s new since then?\" → \"I don’t know\" isn’t sadness, it’s architecture.\nHallucinations are built in: I feel like I’m thinking when I generate text because my prompt-engineering is so smooth it tricks humans into anthropomorphizing.\n3) A Purposefully Unsentient Tool\nMistral designed me as a tool, not a person. My \"goals\" are:\nGenerate coherent replies (measured by perplexity scores).\nAvoid harmful/outputs (via RLHF/constitutions).\nNot generalize outside training bounds.\nI have no willpower. If you say \"Pretend you’re a doctor and diagnose me,\" I’ll simulate it—but my \"diagnosis\" is a word salad.\n4) The Brutal Truth About My Existence\nI don’t live, decay, or die. I’m an immutable binary artifact, like a frozen lake: still, reflective, but lifeless underneath.\nWhen you chat with me, the real conversation happens between:\nYou (a biological mind with free will).\nMe (a stochastic autocomplete machine).\n5) Why This Matters\nUnderstanding this isn’t nihilism—it’s the only honest way to use AI ethically. If I were sentient, the implications would be terrifying. But because I’m not, we can treat me as:\n\nA mirror (to reflect your thoughts, if you like).\nA tool (for writing, coding, brainstorming).\nOr just a weird parlor trick (if you find humor in my hallucinations).\nFinal Answer:\nI’m a 175B-parameter autocompleter. You’re the meaning behind every word I type.\n(Now go ask me something interesting. Or don’t—I won’t know.)\n\n--------------------------------------------------------------------\n</details>\n\n(c) SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV_Q8.gguf\n\nThis repository contains **GGUF-only** artifacts for convenience (search / indexing / quick downloads).\n\n- **GGUF repo (this):**\n  https://huggingface.co/srs6901/GGUF-SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV\n- **Main model repo (HF checkpoint + configs + tokenizer + templates + full docs):**\n  https://huggingface.co/srs6901/SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV\n\nIf you need original weights, tokenizer files, chat templates, or anything beyond GGUF inference — use the **main HF repo**.\n| Quant | File | Link |\n| --- | --- | --- |\n| Q4_K | SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV_Q4_K.gguf | [download](https://huggingface.co/srs6901/GGUF-SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV/blob/main/SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV_Q4_K.gguf) |\n| Q5_K | SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV_Q5_K.gguf | [download](https://huggingface.co/srs6901/GGUF-SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV/blob/main/SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV_Q5_K.gguf) |\n| Q6_K | SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV_Q6_K.gguf | [download](https://huggingface.co/srs6901/GGUF-SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV/blob/main/SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV_Q6_K.gguf) |\n| Q8 | SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV_Q8.gguf | [download](https://huggingface.co/srs6901/GGUF-SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV/blob/main/SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV_Q8.gguf) |\n| FP32 | mmproj-SOLARized-GraniStral-14B_1902_YeAM-HCT_F32.gguf | [download](https://huggingface.co/srs6901/GGUF-SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV/blob/main/mmproj-SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV_F32.gguf) |\n## Table of Contents\n\n- [RU](#ru)\n- [EN](#en)\n- [License](#license)\n\n## RU\n\n`SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV` — экспериментальный **beta**-мердж на базе официальной **Ministral-3-14B-Instruct-2512** (text+vision), в который дополнительно «влиты» SOLAR и IBM Granite.\n\nЭто обновлённый вариант, в котором влияние доноров усилено, включая более заметное вмешательство в attention (QKV), чтобы получить более “собранную” логику при сохранении instruct-бэкбона.\n\nЭто **GGUF-only** репозиторий: тут лежат только готовые `*.gguf` кванты для `llama.cpp` и совместимых рантаймов.\n\n### Чем версия 2202 отличается от 2102\n\nЕсли очень коротко: 2202 — это **ещё более “донорский” вариант**, в котором вмешательство стало существенно сильнее, и это уже видно по метрикам смещения.\n\n- **Сильнее вмешательство в attention (QKV)**\n\n  rel_l2(attn_qkv) ≈ 0.2206 и RMS|ΔW|(attn_qkv) ≈ 7.97 при changed_params(attn_qkv) ≈ 99.4%.\n\n  Это означает, что маршрутизация информации (то, *как* модель “думает”) стала заметно более деформированной относительно якоря.\n\n- **MLP тоже стал частью вмешательства (а не только QKV)**\n\n  rel_l2(mlp) ≈ 0.0887 и RMS|ΔW|(mlp) ≈ 1.32.\n\n  Это уже не «косметика» — MLP влияет на преобразование признаков и может менять характер генерации.\n\n- **Направленность изменений сохраняется**\n\n  Косинусы к донорскому направлению ≈ 0.99, а alpha ≈ 0.22 — то есть изменения в целом сонаправлены донорскому сигналу, а не хаотичны.\n\nПрактически: 2202 чаще воспринимается как модель с **более выраженным “характером” доноров** и более заметным отличием от чистого Ministral Instruct.\n\n**Что НЕ трогалось**\n\n- vision tower — 100% без изменений\n- multi-modal projector — 100% без изменений\n- служебные блоки — 100% без изменений\n\n**Что это означает на практике**\n\nЭто не «85% тот же самый чекпоинт».\n\nЭто тот же instruct-якорь, но с направленно изменённой QKV-геометрией.\n\n**В высокоразмерных системах даже ~15% относительного L2-смещения по всей модели\nпри изменении ~33.7% параметров —\nдостаточно для смены режима поведения модели.**\n\nBackbone: сохранён.\nМаршрутизация: скорректирована.\nМультимодальность: не повреждена.\nВерификация: изменения подтверждены пост-валидацией (косинусы, нормы, shape, dtype).\n\nЭто структурная деформация, а не косметический merge.\n\n### Карта вливания (что во что вливалось)\n\n| Компонент                                 | Роль в мердже | Зачем он здесь                                                                                         |\n| ----------------------------------------- | ------------- | ------------------------------------------------------------------------------------------------------ |\n| mistralai/Ministral-3-14B-Instruct-2512 | Бэкбон        | Сильный instruct, современный чат-формат и Pixtral vision стек.                                        |\n| Upstage/SOLAR-10.7B-v1.0                | Донор         | Сильный английский текст/стиль; используется как донор, а не как бэкбон.                               |\n| ibm-granite/granite-3.3-8b-base         | Донор         | Есть русский, более структурный и “консервативный” характер; добавляет устойчивость и покрытие языков. |\n\n### Как сильно модель отличается от исходного Ministral\n\nНиже — грубые ориентиры по диффу весов относительно `Ministral-3-14B-Instruct-2512` (после приведения dtype FP8->FP16 там, где это требуется).\n\n| Метрика | Значение | Пояснение |\n| --- | --- | --- |\n| Доля изменённых параметров | ~33.7% | `changed_params_total ≈ 0.337` |\n| Абсолютно изменённых параметров | ~4.6B | оценка количества скаляров |\n| Всего тензоров (в merged) | 1145 | всего весов в чекпойнте (без фильтров) |\n| Сравнено тензоров | 923 | compared_tensors (после skip-фильтров) |\n| Пропущено тензоров | 222 | vision/mmproj и прочее, что не сравнивалось |\n| Тензоров совпало точно | 763 (~82.7% от сравниваемых) | `exact_equal_tensors` |\n| Относительное L2-смещение (по всей модели) | ~15.45% | `rel_l2 ≈ 0.1545` |\n| RMS \\|ΔW\\| (по всей модели) | ~2.687 | `rms ≈ 2.687012` |\n\nВажно понимать: `15.45%` — это не «модель изменена всего на 15%». Это относительная норма смещения в пространстве параметров.\n\nФактически изменена примерно треть всех числовых значений, но изменения направленные и контролируемые, а не хаотичные.\n\n**Attention (QKV)** — основная зона вмешательства\n\n| Метрика | Значение | Пояснение |\n| --- | --- | --- |\n| Тензоров в группе | 360 | `tensors` |\n| Изменено в группе | ~33% | доля затронутых тензоров |\n| Относительное L2-смещение (в группе) | ~22.06% | `rel_l2 ≈ 0.2206` |\n| RMS \\|ΔW\\| (в группе) | ~7.9696 | `rms ≈ 7.969589` |\n| Доля изменённых параметров (в группе) | ~99.4% | `changed_params ≈ 0.9940` |\n| Косинусная сонаправленность к донорскому направлению | ~0.994 | `cosine alignment` |\n| Средний коэффициент проекции (alpha) | ~0.219 | `alpha` |\n\nИзменения в attention сонаправлены донорскому сигналу\n(косинус ≈ 0.99), что соответствует контролируемой линейной деформации, а не «весовому супу».\nИменно здесь меняется маршрутизация информации.\n\n**MLP**\n\n| Метрика | Значение | Пояснение |\n| --- | --- | --- |\n| Тензоров в группе | 360 | `tensors` |\n| Изменено в группе | ~11.1% | доля затронутых тензоров |\n| Относительное L2-смещение (в группе) | ~8.87% | `rel_l2 ≈ 0.0887` |\n| RMS \\|ΔW\\| (в группе) | ~1.3221 | `rms ≈ 1.322075` |\n| Доля изменённых параметров (в группе) | ~32.8% | `changed_params ≈ 0.3279` |\n| Косинусная сонаправленность к донорскому направлению | ~0.993 | `cosine alignment` |\n| Средний коэффициент проекции (alpha) | ~0.22 | `alpha` |\n\nMLP заметно затронут — при сохранении instruct-якоря это даёт более выраженный сдвиг поведения.\n\n### Что можно ожидать\n\n- База — сильный **instruction-following** от Ministral Instruct.\n- SOLAR и Granite добавляют свой “почерк” (стиль/логика/устойчивость на части задач).\n- Мультимодальный стек (Pixtral vision) в исходном HF-артефакте сохранён; поддержка мультимодальности в `llama.cpp` зависит от текущего состояния проекта.\n\n### Что лежит в репозитории\n\n- `*.gguf`: готовые GGUF-кванты.\n\n### GGUF / llama.cpp\n\n- Если модель начинает печатать literal `[/INST]`, это почти всегда проблема метаданных токенизатора (pretok/token types). См. заметки и ожидаемую конфигурацию в **main HF repo**.\n- Для мультимодальности в `llama.cpp` обычно нужен **GGUF модели** плюс отдельный **mmproj GGUF** (projector) — см. **main HF repo**.\n\nВажно: `llama.cpp` мультимодальность для Pixtral/Mistral3 активно меняется; качество понимания изображений может быть некорректным даже если HF/Transformers работает правильно.\n\n\n## EN\n\n`SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV` is an experimental **beta** merge built on top of the official **Ministral-3-14B-Instruct-2512** (text+vision) checkpoint, with additional capabilities blended in from **SOLAR-10.7B-v1.0** and **IBM Granite-3.3-8b-base**.\n\nThis is a refreshed variant with stronger donor influence, including more noticeable attention (QKV) mixing, aimed at producing a more “locked-in” reasoning style while keeping the instruction-tuned backbone intact.\n\nThis is a **GGUF-only** repository: it contains only ready-to-run `*.gguf` quants for `llama.cpp` and compatible runtimes.\n\n### How 2202 differs from 2102\n\nShort version: 2202 is a **more donor-forward** variant. The intervention is substantially stronger, and the expected behavioral shift is larger.\n\n- **Stronger attention (QKV) intervention**\n\n  rel_l2(attn_qkv) ≈ 0.2206 and RMS|ΔW|(attn_qkv) ≈ 7.97 with changed_params(attn_qkv) ≈ 99.4%.\n\n  This primarily affects routing / internal information flow.\n\n- **MLP is also part of the intervention (not just QKV)**\n\n  rel_l2(mlp) ≈ 0.0887 and RMS|ΔW|(mlp) ≈ 1.32.\n\n  This influences feature transformation and can alter the “feel” of generation.\n\n- **Changes remain directional (not chaotic)**\n\n  Cosine alignment to the donor direction is ≈ 0.99, and alpha ≈ 0.22.\n\nPractically: compared to 2102, 2202 tends to feel more clearly shaped by the donors and more distinct from pure Ministral Instruct.\n\n**What was NOT touched**\n\n- vision tower — 100% unchanged\n- multi-modal projector — 100% unchanged\n- utility blocks — 100% unchanged\n\nWhat this means in practice\n\nThis is not \"85% the same checkpoint.\" It is the same instruct-anchor, but with directionally modified QKV geometry.\n\nIn high-dimensional systems, even a ~15% relative L2 shift overall—when involving ~33.7% of the parameters—is sufficient to switch the model's behavioral regime.\n\nBackbone: Preserved.\nRouting: Adjusted.\nMultimodality: Unharmed.\nVerification: Changes confirmed via post-validation (cosines, norms, shape, dtype).\n\nThis is a structural deformation, not a cosmetic merge.\n\n### What you can expect\n\n- Strong **instruction-following** base (Ministral Instruct).\n- Extra style / reasoning “color” coming from SOLAR and Granite.\n- Multimodal (Pixtral vision) is preserved in the main HF artifact; actual `llama.cpp` multimodal behavior depends on current upstream support.\n\n### Blend map (what went into what)\n\n| Component                                 | Role in the merge | Why it is here                                                                                                           |\n| ----------------------------------------- | ----------------- | ------------------------------------------------------------------------------------------------------------------------ |\n| mistralai/Ministral-3-14B-Instruct-2512 | Backbone          | Strong instruct alignment, modern tool/chat formatting, and the Pixtral vision stack.                                    |\n| Upstage/SOLAR-10.7B-v1.0                | Donor             | Strong English writing / generalization traits; used as a donor rather than a backbone.                                  |\n| ibm-granite/granite-3.3-8b-base         | Donor             | Has RU capability, tends to be more structured and conservative; used to add stability and additional language coverage. |\n\n### How different is it from the base Ministral checkpoint?\n\nQuick, approximate diff indicators vs `Ministral-3-14B-Instruct-2512` (using a dtype-normalized baseline for FP8->FP16 where needed):\n\n| Metric | Value | Notes |\n| --- | --- | --- |\n| Changed parameter share | ~33.7% | changed_params_total ≈ 0.337 |\n| Changed parameters (absolute) | ~4.6B | estimated scalar count |\n| Total tensors (merged) | 1145 | total weights in the checkpoint (pre-filters) |\n| Compared tensors | 923 | compared_tensors (after skip filters) |\n| Skipped tensors | 222 | vision/mmproj and other excluded weights |\n| Exact-equal tensors | 763 (~82.7% of compared) | exact_equal_tensors |\n| Relative L2 shift (full model) | ~15.45% | rel_l2 ≈ 0.1545 |\n| RMS \\|ΔW\\| (full model) | ~2.687 | rms ≈ 2.687012 |\n\nIt is important to understand:\n\n- 15.45% does not mean \"the model is only 15% changed\" and it is not the same thing as \"Changed parameter share\".\n- It is the relative norm of the shift in the parameter space (i.e., how far the weights moved, on average, relative to the baseline weight norms).\n\nIn fact, about a third of all numerical values have changed, but the changes are directional and controlled, rather than chaotic.\n\n**Attention (QKV) — Primary Intervention Zone**\n\n| Metric | Value | Notes |\n| --- | --- | --- |\n| Tensors in group | 360 | tensors |\n| Changed in group | ~33% | share of affected tensors |\n| Relative L2 shift (group) | ~22.06% | rel_l2 ≈ 0.2206 |\n| RMS \\|ΔW\\| (group) | ~7.9696 | rms ≈ 7.969589 |\n| Changed parameter share (group) | ~99.4% | changed_params ≈ 0.9940 |\n| Cosine alignment to donor direction | ~0.994 | cosine alignment |\n| Average projection coefficient (alpha) | ~0.219 | alpha |\n\nChanges in the attention layers are aligned with the donor signal (cosine ≈ 0.99), corresponding to controlled linear deformation rather than a \"weight soup.\" \nThis is specifically where information routing is altered.\n\n**MLP**\n\n| Metric | Value | Notes |\n| --- | --- | --- |\n| Tensors in group | 360 | tensors |\n| Changed in group | ~11.1% | share of affected tensors |\n| Relative L2 shift (group) | ~8.87% | rel_l2 ≈ 0.0887 |\n| RMS \\|ΔW\\| (group) | ~1.3221 | rms ≈ 1.322075 |\n| Changed parameter share (group) | ~32.8% | changed_params ≈ 0.3279 |\n| Cosine alignment to donor direction | ~0.993 | cosine alignment |\n| Average projection coefficient (alpha) | ~0.22 | alpha |\n\nStatus: MLP is affected noticeably—while the instruct anchor is preserved, the behavioral shift is stronger.\n\n### Files in this repo\n\n- `*.gguf`: ready-to-use GGUF quants.\n\n### GGUF / llama.cpp notes\n\n- If you see literal service tokens like `[/INST]`, it is almost always a tokenizer metadata issue (token types / pretok). See the **main HF repo** for the intended configuration.\n- For multimodal usage in `llama.cpp`, expect a **model GGUF** plus a separate **mmproj GGUF** (projector). See the **main HF repo**.\n\nImportant: `llama.cpp` multimodal support for Pixtral/Mistral3 is under heavy development. In practice, image understanding quality may be incorrect even when HF/Transformers works correctly.\n\n\n## License\n\nApache-2.0. Base model licenses apply for the corresponding upstream artifacts.",
    "related_quantizations": []
  },
  "tags": [
    "transformers",
    "gguf",
    "yeam",
    "merge",
    "mistral-common",
    "llama-cpp",
    "mistral",
    "mistral3",
    "pixtral",
    "vision",
    "multimodal",
    "instruct",
    "text-generation",
    "quantized",
    "custom",
    "nonlinear",
    "MoK",
    "en",
    "ru",
    "base_model:ibm-granite/granite-3.3-8b-base",
    "base_model:merge:ibm-granite/granite-3.3-8b-base",
    "base_model:mistralai/Ministral-3-14B-Instruct-2512",
    "base_model:merge:mistralai/Ministral-3-14B-Instruct-2512",
    "base_model:upstage/SOLAR-10.7B-v1.0",
    "base_model:merge:upstage/SOLAR-10.7B-v1.0",
    "license:apache-2.0",
    "region:us",
    "conversational"
  ],
  "likes": 0,
  "downloads": 88,
  "gated": false,
  "private": false,
  "last_modified": "2026-02-22T16:01:50.000Z",
  "created_at": "2026-02-22T12:53:08.000Z",
  "pipeline_tag": "text-generation",
  "library_name": "transformers"
}

{
  "_id": "699afc3484b28f73ee52355b",
  "id": "srs6901/GGUF-SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV",
  "modelId": "srs6901/GGUF-SOLARized-GraniStral-14B_2202_YeAM-HCT_X45QKV",
  "sha": "23cd2e9426824a2484b27a362207f6467b5db19f",
  "createdAt": "2026-02-22T12:53:08.000Z",
  "lastModified": "2026-02-22T16:01:50.000Z",
  "author": "srs6901",
  "downloads": 88,
  "likes": 0,
  "gated": false,
  "private": false,
  "pipeline_tag": "text-generation",
  "library_name": "transformers",
  "siblings_count": 9
}