AutoPowerPoint Agent — Golden v2 Cycle 33
This is the merged Hugging Face release build for AutoPowerPoint Agent. It combines the unsloth/Qwen3.5-2B base model with the best Reinforcement SFT checkpoint selected by the Golden v2 and challenge promotion gates.
Selection metrics
| Metric | Score |
|---|---|
| Runtime success | 85% |
| Layout success | 67% |
| Studio success | 14% |
| Content success | 64% |
| Strict Studio success | 10% |
| Challenge Strict Studio success | 20% |
Cycle 33 was promoted because it remained within the Golden regression tolerance and improved the challenge and combined promotion scores over the previous champion.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "YOUR_ORG/autopowerpoint-agent-golden-v2-cycle33"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
torch_dtype="auto",
trust_remote_code=True,
)
The model is intended for local PowerPoint planning, slide-code generation, and repair workflows. Generated Python must still pass compile, static, runtime, PPTX, and layout validation before use.
Build provenance
- Reinforcement job:
4622f449-63d7-4305-bdf8-808fc82696ce - Champion cycle:
33 - Base model:
unsloth/Qwen3.5-2B - Export dtype:
bfloat16 - Format: merged
safetensors