Downloads · 30 days
5
11% of all-time downloads
qwen3.8-27b-grpo-dpo
Knowlify/qwen3.8-27b-grpo-dpo
qwen3.8-27b-grpo-dpo is a machine learning model from Knowlify. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Qwen3.8-27B stacked SFT → GRPO (best) → DPO for Remotion scene generation.
All-time downloads
46
Public
Parameters
26.9B
54.4 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors54.4 GB · 100%
At a glance
- License
- apache-2.0
- Model type
- qwen3_5_text
- Access
- Public
- Created
- Aug 29, 2026
- Updated
- Aug 29, 2026
- SHA
- edc8c964
Base models
Model card
From the Hugging Face model README
Knowlify/qwen3.8-27b-grpo-dpo
Qwen3.8-27B stacked SFT → GRPO (best) → DPO for Remotion scene generation.
Pipeline:
- SFT:
Knowlify/qwen3.8-27b-sft-ge75 - GRPO on SFT (gated reward: compile/render gate + qwen3.8-max video judge), best checkpoint (step 25) merged onto SFT
- DPO on that merge:
Knowlify/fablify-remotion-dpopairs, score margin ≥ 15, max seq 6144, 1 epoch (1341 pairs / 168 steps), LoRA r32
Final train metrics: preference accuracy ~0.875-1.0, margins +2.0 to +3.5 (from 0 at step 0), loss 0.69 → ~0.25.
Repo layout: merged 16-bit weights at root, DPO LoRA adapter under lora/.
Video-judge eval vs the DPO and SFT baselines is pending; treat as experimental until graded.
- Type
- qwen3_5_text
- License
- apache-2.0
- Created
- Aug 29, 2026
- Updated
- Aug 29, 2026