43 lines
2.6 KiB
JSON
43 lines
2.6 KiB
JSON
{
|
|
"schema": "fable.research_map.receipt.v1",
|
|
"createdAt": "2026-07-05T01:35:00.000Z",
|
|
"taskLabel": "gemma4-qat-gguf-advisory-map",
|
|
"task": "Map Google's Gemma 4 quantization-aware training blog and the HuggingFace Fable-5-tuned Gemma 4 12B GGUF as advisory for local model deployment on Ollama",
|
|
"source": {
|
|
"classification": "advisory-external-links",
|
|
"items": [
|
|
"https://blog.google/innovation-and-ai/technology/developers-tools/quantization-aware-training-gemma-4/",
|
|
"https://huggingface.co/yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF"
|
|
],
|
|
"contentVerified": false,
|
|
"localProofExists": false,
|
|
"liveVerification": "not fetched — links provided by user during Ollama setup session"
|
|
},
|
|
"extractedPrinciples": [
|
|
"Gemma 4 uses quantization-aware training (QAT) — model was trained knowing it would be quantized, so quantized versions lose less quality than post-hoc quantization",
|
|
"Google claims 4-bit QAT Gemma 4 approaches FP16 quality — significant for running on consumer/VPS hardware",
|
|
"The HuggingFace model is a Fable 5 agentic composer fine-tune of Gemma 4 12B (not 27B), converted to GGUF format",
|
|
"GGUF format is directly compatible with Ollama's llama.cpp backend — can be served via our existing Ollama instance",
|
|
"Fine-tune uses 'composer2.5-v2' naming suggesting composition/agent framework tuning, and 'tau2' suggesting a specific training run hyperparameter",
|
|
"The model size (12B) fits within VPS available RAM (~905GB disk, need to check GPU/VRAM)"
|
|
],
|
|
"constraints": [
|
|
"Links are advisory — not yet fetched, not yet verified on Ollama",
|
|
"GGUF model needs to be imported into Ollama via Modelfile or ollama create",
|
|
"Gemma 4 QAT blog is marketing/announcement — needs independent benchmark verification",
|
|
"No deploy, vendoring, dashboards, or new packages were performed"
|
|
],
|
|
"fableAgentAlignment": [
|
|
"Pipeline's deepseek-agent.ts already supports configurable endpoint/model — can target gemma4:12b by changing model name",
|
|
"RunAgent callback takes model config, so switching from deepseek-coder-v2 to gemma4 is a one-line change",
|
|
"If Gemma 4 QAT delivers on its quality claims, it's the best candidate for the pipeline's agent runtime: smaller, faster, free, local"
|
|
],
|
|
"skipped": [
|
|
"actual model download/pull on Ollama (waiting on deepseek-coder-v2 to finish at 86%)",
|
|
"benchmark comparisons between deepseek-coder-v2 and gemma4 on coding tasks"
|
|
],
|
|
"decision": "advisory — explore after deepseek-coder-v2 finishes pulling",
|
|
"deployAttempted": false,
|
|
"autoPatchProven": false
|
|
}
|