bigatuna/qwen3.5-9b-sushi-coder-rl-gguf overview

{
  "metadata": {},
  "card_data": {
    "license": "apache-2.0",
    "language": [
      "en"
    ],
    "pipeline_tag": "text-generation",
    "base_model": "bigatuna/Qwen3.5-9b-Sushi-Coder-RL",
    "base_model_relation": "quantized",
    "datasets": [
      "open-r1/codeforces-cots",
      "nohurry/Opus-4.6-Reasoning-3000x-filtered"
    ],
    "tags": [
      "gguf",
      "llama.cpp",
      "qwen3_5",
      "code",
      "rl",
      "atropos",
      "multimodal"
    ],
    "frontmatter": {
      "license": "apache-2.0",
      "language": [
        "en"
      ],
      "pipeline_tag": "text-generation",
      "base_model": "bigatuna/Qwen3.5-9b-Sushi-Coder-RL",
      "base_model_relation": "quantized",
      "datasets": [
        "open-r1/codeforces-cots",
        "nohurry/Opus-4.6-Reasoning-3000x-filtered"
      ],
      "tags": [
        "gguf",
        "llama.cpp",
        "qwen3_5",
        "code",
        "rl",
        "atropos",
        "multimodal"
      ]
    },
    "hero_image_url": "./qwen3.5-9b.png",
    "summary": "!Qwen3.5-9b-Sushi-Coder-RL-GGUF",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlicense: apache-2.0\nlanguage:\n- en\npipeline_tag: text-generation\nbase_model: bigatuna/Qwen3.5-9b-Sushi-Coder-RL\nbase_model_relation: quantized\ndatasets:\n- open-r1/codeforces-cots\n- nohurry/Opus-4.6-Reasoning-3000x-filtered\ntags:\n- gguf\n- llama.cpp\n- qwen3_5\n- code\n- rl\n- atropos\n- multimodal\n---\n\n# Qwen3.5-9b-Sushi-Coder-RL-GGUF\n\n![Qwen3.5-9b-Sushi-Coder-RL-GGUF](./qwen3.5-9b.png)\n\n## Lineage\n\n- Base model lineage: [`bigatuna/Qwen3.5-9b-Sushi-Coder`](https://huggingface.co/bigatuna/Qwen3.5-9b-Sushi-Coder)\n- RL model: [`bigatuna/Qwen3.5-9b-Sushi-Coder-RL`](https://huggingface.co/bigatuna/Qwen3.5-9b-Sushi-Coder-RL)\n- RL pipeline: [NousResearch/atropos](https://github.com/NousResearch/atropos)\n\n## Training\n\nThe upstream SFT model was trained with Unsloth on:\n\n- [`nohurry/Opus-4.6-Reasoning-3000x-filtered`](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered)\n- [`open-r1/codeforces-cots`](https://huggingface.co/datasets/open-r1/codeforces-cots)\n\nThe RL stage was then run for coding with [NousResearch/hermes-agent](https://github.com/NousResearch/hermes-agent) using [NousResearch/atropos](https://github.com/NousResearch/atropos).\n\nDuring that run, vLLM was patched with [`vllm-project/vllm` PR #36395](https://github.com/vllm-project/vllm/pull/36395), `fix(lora): add bounds checking for TP configurations`, to address the LoRA tensor-parallel bounds issue.\n\n## Files\n\n- `Qwen3.5-9b-Sushi-Coder-RL.Q4_K_M.gguf`\n- `Qwen3.5-9b-Sushi-Coder-RL.Q8_0.gguf`\n- `Qwen3.5-9b-Sushi-Coder-RL.BF16-mmproj.gguf`\n\n## Usage Note\n\nThis is a multimodal Qwen 3.5 export. Use the text GGUF together with the `BF16-mmproj` file.\n\n## Quick Start\n\nExample download commands with the Hugging Face CLI:\n\n```bash\nhf download bigatuna/Qwen3.5-9b-Sushi-Coder-RL-GGUF \\\n  Qwen3.5-9b-Sushi-Coder-RL.Q4_K_M.gguf \\\n  Qwen3.5-9b-Sushi-Coder-RL.BF16-mmproj.gguf\n```\n\nAlternative quant:\n\n```bash\nhf download bigatuna/Qwen3.5-9b-Sushi-Coder-RL-GGUF \\\n  Qwen3.5-9b-Sushi-Coder-RL.Q8_0.gguf \\\n  Qwen3.5-9b-Sushi-Coder-RL.BF16-mmproj.gguf\n```\n\n## Metadata\n\n- License: Apache-2.0\n- Architecture: Qwen 3.5\n- Format: GGUF\n- Tags: `llama.cpp`, `qwen3_5`, `multimodal`, `code`, `rl`, `conversational`\n\n## LiveCodeBench Evaluation\n\nThe benchmark results below were produced from matched local BF16 vLLM endpoints so the RL model and the base model were evaluated with the same serving method, the same task, and the same generation settings.\n\nEvaluated models:\n\n- RL model: `bigatuna/Qwen3.5-9b-Sushi-Coder-RL`\n- Base model: `Qwen/Qwen3.5-9B`\n\nMatched benchmark setup:\n\n- Task: `lcb:codegeneration|0`\n- Benchmark size: `268` problems\n- Backend: `lighteval endpoint litellm`\n- Context length: `4096`\n- vLLM dtype: `bfloat16`\n- Same `max_new_tokens`, same prompt/task, same serving stack, same evaluation harness\n\n### Matched Full Results\n\nDeterministic run:\n\n- `temperature=0.0`\n- `top_p=1.0`\n- `seed=0`\n- `max_new_tokens=1024`\n\nResults:\n\n- RL model: `codegen_pass@1:16 = 0.2015 +/- 0.0245`\n- Base model: `codegen_pass@1:16 = 0.0336 +/- 0.0110`\n\nApproximate passes:\n\n- RL model: `54 / 268`\n- Base model: `9 / 268`\n\nSampling run:\n\n- `temperature=0.6`\n- `top_p=0.95`\n- `top_k=20`\n- `min_p=0.0`\n- `presence_penalty=0.0`\n- `repetition_penalty=1.0`\n- `max_new_tokens=1024`\n\nResults:\n\n- RL model: `codegen_pass@1:16 = 0.2388 +/- 0.0261`\n- Base model: `codegen_pass@1:16 = 0.0261 +/- 0.0098`\n\nApproximate passes:\n\n- RL model: `64 / 268`\n- Base model: `7 / 268`\n\nIn both matched full runs, the RL model outperformed the base model by a wide margin.\n\n## Exact Reproduction Commands\n\nThese are the exact command shapes used for the matched local evaluation.\n\n### 1. Start the RL endpoint\n\n```bash\nexport CUDA_VISIBLE_DEVICES=1\nvllm serve \\\n  <PATH_TO_YOUR_RL_MERGED_MODEL> \\\n  --host 0.0.0.0 \\\n  --port 9001 \\\n  --served-model-name bigatuna/Qwen3.5-9b-Sushi-Coder-RL \\\n  --max-model-len 4096 \\\n  --dtype bfloat16 \\\n  --gpu-memory-utilization 0.45\n```\n\n### 2. Start the base endpoint\n\n```bash\nexport CUDA_VISIBLE_DEVICES=0\nvllm serve \\\n  Qwen/Qwen3.5-9B \\\n  --host 0.0.0.0 \\\n  --port 9002 \\\n  --served-model-name Qwen/Qwen3.5-9B \\\n  --max-model-len 4096 \\\n  --dtype bfloat16 \\\n  --gpu-memory-utilization 0.45\n```\n\n### 3. Deterministic matched run\n\nRL model:\n\n```bash\ncat > /tmp/lighteval_rl.yaml <<'EOF'\nmodel_parameters:\n  provider: \"openai\"\n  model_name: \"openai/bigatuna/Qwen3.5-9b-Sushi-Coder-RL\"\n  base_url: \"http://localhost:9001/v1\"\n  api_key: \"dummy\"\n  generation_parameters:\n    temperature: 0.0\n    max_new_tokens: 1024\n    top_p: 1.0\n    seed: 0\nEOF\n\nlighteval endpoint litellm \\\n  /tmp/lighteval_rl.yaml \\\n  'lcb:codegeneration|0' \\\n  --output-dir /tmp/lcb_rl_full \\\n  --save-details\n```\n\nBase model:\n\n```bash\ncat > /tmp/lighteval_base.yaml <<'EOF'\nmodel_parameters:\n  provider: \"openai\"\n  model_name: \"openai/Qwen/Qwen3.5-9B\"\n  base_url: \"http://localhost:9002/v1\"\n  api_key: \"dummy\"\n  generation_parameters:\n    temperature: 0.0\n    max_new_tokens: 1024\n    top_p: 1.0\n    seed: 0\nEOF\n\nlighteval endpoint litellm \\\n  /tmp/lighteval_base.yaml \\\n  'lcb:codegeneration|0' \\\n  --output-dir /tmp/lcb_base_full \\\n  --save-details\n```\n\n### 4. Temperature 0.6 matched run\n\nRL model:\n\n```bash\ncat > /tmp/lighteval_rl_t06.yaml <<'EOF'\nmodel_parameters:\n  provider: \"openai\"\n  model_name: \"openai/bigatuna/Qwen3.5-9b-Sushi-Coder-RL\"\n  base_url: \"http://localhost:9001/v1\"\n  api_key: \"dummy\"\n  generation_parameters:\n    temperature: 0.6\n    max_new_tokens: 1024\n    top_p: 0.95\n    top_k: 20\n    min_p: 0.0\n    presence_penalty: 0.0\n    repetition_penalty: 1.0\nEOF\n\nlighteval endpoint litellm \\\n  /tmp/lighteval_rl_t06.yaml \\\n  'lcb:codegeneration|0' \\\n  --output-dir /tmp/lcb_rl_full_t06 \\\n  --save-details\n```\n\nBase model:\n\n```bash\ncat > /tmp/lighteval_base_t06.yaml <<'EOF'\nmodel_parameters:\n  provider: \"openai\"\n  model_name: \"openai/Qwen/Qwen3.5-9B\"\n  base_url: \"http://localhost:9002/v1\"\n  api_key: \"dummy\"\n  generation_parameters:\n    temperature: 0.6\n    max_new_tokens: 1024\n    top_p: 0.95\n    top_k: 20\n    min_p: 0.0\n    presence_penalty: 0.0\n    repetition_penalty: 1.0\nEOF\n\nlighteval endpoint litellm \\\n  /tmp/lighteval_base_t06.yaml \\\n  'lcb:codegeneration|0' \\\n  --output-dir /tmp/lcb_base_full_t06 \\\n  --save-details\n```\n",
    "related_quantizations": []
  },
  "tags": [
    "gguf",
    "llama.cpp",
    "qwen3_5",
    "code",
    "rl",
    "atropos",
    "multimodal",
    "text-generation",
    "en",
    "dataset:open-r1/codeforces-cots",
    "dataset:nohurry/Opus-4.6-Reasoning-3000x-filtered",
    "base_model:bigatuna/Qwen3.5-9b-Sushi-Coder-RL",
    "base_model:quantized:bigatuna/Qwen3.5-9b-Sushi-Coder-RL",
    "license:apache-2.0",
    "endpoints_compatible",
    "region:us",
    "conversational"
  ],
  "likes": 50,
  "downloads": 19155,
  "gated": false,
  "private": false,
  "last_modified": "2026-03-31T19:07:44.000Z",
  "created_at": "2026-03-27T01:12:29.000Z",
  "pipeline_tag": "text-generation",
  "library_name": ""
}

{
  "_id": "69c5d97d23056adf63dc6cd4",
  "id": "bigatuna/Qwen3.5-9b-Sushi-Coder-RL-GGUF",
  "modelId": "bigatuna/Qwen3.5-9b-Sushi-Coder-RL-GGUF",
  "sha": "36816bdeb384c86c786847f6a5a779594012ce1f",
  "createdAt": "2026-03-27T01:12:29.000Z",
  "lastModified": "2026-03-31T19:07:44.000Z",
  "author": "bigatuna",
  "downloads": 19155,
  "likes": 50,
  "gated": false,
  "private": false,
  "pipeline_tag": "text-generation",
  "library_name": "",
  "siblings_count": 8
}