Downloads · 30 days
6
8% of all-time downloads
qwen3.8-27b-dpo-grpo
Knowlify/qwen3.8-27b-dpo-grpo
qwen3.8-27b-dpo-grpo 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 → DPO → GRPO for Remotion scene generation.
All-time downloads
75
Public
Parameters
26.9B
54.5 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 30, 2026
- Updated
- Aug 30, 2026
- SHA
- f13d6f44
Base models
Model card
From the Hugging Face model README
Knowlify/qwen3.8-27b-dpo-grpo
Qwen3.8-27B stacked SFT → DPO → GRPO for Remotion scene generation.
Pipeline:
- SFT:
Knowlify/qwen3.8-27b-sft-ge75 - DPO on SFT:
Knowlify/fablify-remotion-dpopairs, margin >= 15, seq 8192, 1 epoch ->Knowlify/qwen3.8-27b-dpo - GRPO on that DPO merge: gated reward (compile/render/layout gate + qwen3.8-max video judge: content 25 / layout 20 / motion 20 / polish 15 / color 10 / type 10), KL beta 0.03, LoRA r32, 100 steps over
Knowlify/fablify-remotion-synth-prompts
GRPO training curve (quality = gated judge score): 2.21 (steps 1-25) -> 4.01 (26-50) -> 5.63 (51-75) -> 6.51 (76-100), compile gate 72% -> 96%, KL 0.001 -> 0.018, no collapse. Best single step 98 (quality 8.28).
Repo layout: merged 16-bit weights at root, final GRPO LoRA adapter under lora/. Intermediate GRPO checkpoints (25/50/75/100) were not uploaded; they remain on the training box until it is released.
Video-judge eval vs Knowlify/qwen3.8-27b-dpo is pending; final-step adapter shipped pending A/B grading of the intermediate checkpoints.
- Type
- qwen3_5_text
- License
- apache-2.0
- Created
- Aug 30, 2026
- Updated
- Aug 30, 2026