Article
Rewarding fact fidelity with an LLM judge in GRPO: lessons from 38,000 judged rewrites
jialinyyzz
• • 3
Evan
Gtmix
·
AI & ML interests
None yet
Recent Activity
upvoted an article 16 days ago
Rewarding fact fidelity with an LLM judge in GRPO: lessons from 38,000 judged rewrites liked a Space 16 days ago
jialinyyzz/humanizer liked a model 16 days ago
LessThanThreeAI/Qwen3.8-27B-Humanlike-Chat-GGUFOrganizations
upvoted an article 16 days ago
liked a
Space 16 days ago
liked a
model 16 days ago
liked a
dataset 28 days ago
upvoted a changelog 3 months ago
Hugging Face Changelog
MCP Server Enhancements
• 252
PROJECT IS ACTIVE AGAIN ✨ IMPROVED FOR FREE & PREMIUM USERS 🥳
❤️ 2
1
#51 opened 6 months ago
by 
NihalGazi
liked a
dataset 7 months ago
reacted to qgallouedec's post with 🔥 7 months ago
Post
2,906
TRL v1.2 introduces the SSDTrainer 🚀
Simple Self-Distillation (SSD) from Apple's paper "Embarrassingly Simple Self-Distillation Improves Code Generation" is now available as an experimental trainer in TRL.
The recipe is as minimal as the name suggests: sample completions from the model itself at a training-time temperature, then fine-tune on those raw, unverified samples with plain cross-entropy. No reward model. No verifier. No teacher model. No reinforcement learning. Just prompts and the model.
v1.2 also ships expanded tool-calling support (LLaMA 3.1 / 3.2, DeepSeek-V3), another round of KTO ↔ DPO alignment getting us closer to promoting KTO to stable, a big GRPO simplification for overlong tool results, deprecation of
Full release notes: https://github.com/huggingface/trl/releases/tag/v1.2.0
Simple Self-Distillation (SSD) from Apple's paper "Embarrassingly Simple Self-Distillation Improves Code Generation" is now available as an experimental trainer in TRL.
The recipe is as minimal as the name suggests: sample completions from the model itself at a training-time temperature, then fine-tune on those raw, unverified samples with plain cross-entropy. No reward model. No verifier. No teacher model. No reinforcement learning. Just prompts and the model.
from trl.experimental.ssd import SSDConfig, SSDTrainer
trainer = SSDTrainer(
model="Qwen/Qwen3-4B-Instruct",
args=SSDConfig(temperature=0.6, top_k=20, top_p=0.95),
train_dataset=dataset,
)
trainer.train()v1.2 also ships expanded tool-calling support (LLaMA 3.1 / 3.2, DeepSeek-V3), another round of KTO ↔ DPO alignment getting us closer to promoting KTO to stable, a big GRPO simplification for overlong tool results, deprecation of
use_transformers_paged, and key fixes for VLM response parsing.Full release notes: https://github.com/huggingface/trl/releases/tag/v1.2.0
liked a
Space 7 months ago
liked a
model 7 months ago
liked 2
Spaces 8 months ago
liked a
dataset 8 months ago
liked 7
Spaces 8 months ago
KoboldAI Lite
👁
78
Generate AI-powered story continuations
TADA
🎵
88
Generate speech that clones voice, emotion, and timing from a prompt
VividFlow
🌊
41
Bring your images to life
DeepSite v4
🐳
16.6k
Generate any application by Vibe Coding it
cbc
🐳
1
FFmpeg
🔥
9
Convert media files to different formats
Nymbo Tools MCP
⚙
59
All-in-one hub of general purpose tools useful for any agent