初始化项目，由ModelHub XC社区提供模型

Model: Omarrran/koshur-kouter-ks-en_v1
Source: Original Platform

notebooks/colab_load_stage1_model.ipynb

@@ -0,0 +1,129 @@
+{
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "# Koshur Kouter v1 — Colab loader\n",
+        "\n",
+        "This notebook loads the **full merged Stage 1** model from Hugging Face and runs a few example translations.\n",
+        "\n",
+        "- Repo: `Omarrran/koshur-kouter-ks-en_v1`\n",
+        "- Base model family: `sarvamai/sarvam-translate` / Gemma 3 text stack\n",
+        "- Recommended for full `bfloat16` loading on Colab: **L4 / A100 / better**\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "!pip -q install -U transformers accelerate sentencepiece safetensors\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "import torch\n",
+        "from transformers import AutoTokenizer, AutoModelForCausalLM\n",
+        "\n",
+        "REPO_ID = \"Omarrran/koshur-kouter-ks-en_v1\"\n",
+        "\n",
+        "tokenizer = AutoTokenizer.from_pretrained(REPO_ID)\n",
+        "model = AutoModelForCausalLM.from_pretrained(\n",
+        "    REPO_ID,\n",
+        "    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,\n",
+        "    device_map=\"auto\",\n",
+        ")\n",
+        "model.eval()\n",
+        "print(model.__class__.__name__)\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "def first_nonempty_line(text: str) -> str:\n",
+        "    for line in text.splitlines():\n",
+        "        line = line.strip()\n",
+        "        if line:\n",
+        "            return line\n",
+        "    return text.strip()\n",
+        "\n",
+        "def build_prompt(source_text: str, direction: str):\n",
+        "    if direction == \"ks2en\":\n",
+        "        system = \"Translate the text below to English. Return only the translation.\"\n",
+        "    elif direction == \"en2ks\":\n",
+        "        system = \"Translate the text below to Kashmiri. Return only the translation.\"\n",
+        "    else:\n",
+        "        system = \"Translate the text below. Return only the translation.\"\n",
+        "    return [\n",
+        "        {\"role\": \"system\", \"content\": system},\n",
+        "        {\"role\": \"user\", \"content\": source_text},\n",
+        "    ]\n",
+        "\n",
+        "def translate(source_text: str, direction: str, max_new_tokens: int = 48):\n",
+        "    messages = build_prompt(source_text, direction)\n",
+        "    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
+        "    inputs = tokenizer(prompt, return_tensors=\"pt\", truncation=True, max_length=1024).to(model.device)\n",
+        "    with torch.no_grad():\n",
+        "        outputs = model.generate(\n",
+        "            **inputs,\n",
+        "            max_new_tokens=max_new_tokens,\n",
+        "            do_sample=False,\n",
+        "            repetition_penalty=1.15,\n",
+        "            no_repeat_ngram_size=3,\n",
+        "            pad_token_id=tokenizer.eos_token_id,\n",
+        "        )\n",
+        "    decoded = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)\n",
+        "    return first_nonempty_line(decoded)\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "examples = [\n",
+        "    (\"Please sit down and have some tea.\", \"en2ks\"),\n",
+        "    (\"The children are playing in the garden.\", \"en2ks\"),\n",
+        "    (\"کٔشیر چھُ اکھ خوبصورت جٲی\", \"ks2en\"),\n",
+        "]\n",
+        "\n",
+        "for text, direction in examples:\n",
+        "    print(\"SOURCE:\", text)\n",
+        "    print(\"DIRECTION:\", direction)\n",
+        "    print(\"PRED:\", translate(text, direction))\n",
+        "    print(\"-\" * 80)\n"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## Optional lower-memory path\n",
+        "\n",
+        "If your Colab GPU is tight on VRAM, you can try a quantized load with `bitsandbytes`. The uploaded repository is still a **full merged model**; quantized loading is only a runtime option.\n"
+      ]
+    }
+  ],
+  "metadata": {
+    "kernelspec": {
+      "display_name": "Python 3",
+      "language": "python",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 5
+}