Activity Composition Model 400 — GGUF

Quantized GGUF builds of mjpsm/activity-composition-model-400-qwen3.5-0.8b for MyVillage Activity Composition inference.

Input

The model receives only the Activity Progression Model output.

Output

{"output":{"activity_title":"...","activity_description":"..."}}

Quantization candidates

• activity-composition-model-400-qwen3.5-0.8b-Q8_0.gguf
  • Recommended first API deployment candidate.
• activity-composition-model-400-qwen3.5-0.8b-Q4_K_M.gguf
  • Smaller experimental candidate; use only if regression testing shows acceptable behavior.

Source

The source checkpoint is the already-merged model:

mjpsm/activity-composition-model-400-qwen3.5-0.8b

Deployment

The intended deployment target is a Hugging Face Docker Space using llama.cpp.

Validation

The quantized candidates should be compared against the F16 GGUF baseline on the fixed 20-case unseen Composition challenge set before selecting the production artifact.