Meanblock/JEV-CPU
Zero-Shot Classification ā¢ Updated ā¢ 41
eoe
eoe
AI & ML interests
None yet
Recent Activity
liked a model 3 days ago
Meanblock/JEV-CPU liked a model 29 days ago
Qwen/Qwen3.8-Flash-Next upvoted a collection 5 months ago
2026 April š China Open Source Highlights Organizations
None yet
liked a
model 3 days ago
liked a
model 29 days ago
upvoted a collection 5 months ago
reacted to anakin87's post with ā¤ļø 5 months ago
Post
10423
How LLM training with RL Environments works?
It all starts with š„š²š¶š»š³š¼šæš°š²šŗš²š»š šš²š®šæš»š¶š»š“ šš¶ššµ š©š²šæš¶š³š¶š®šÆš¹š² š„š²šš®šæš±š
- question asked
- model generates reasoning + answer
- answer checked against ground truth
- reward drives RL training
In this setup, the environment is simple: fixed questions and answers, rollout logic, reward(s)
Consider a more complex tic-tac-toe env āā
It adds:
- dynamic game generation/handling
- tunable opponent skill
- multi-turn interactions
(envs can also include tools)
---
What happens at training?
We use ššæš¼šš½ š„š²š¹š®šš¶šš² š£š¼š¹š¶š°š š¢š½šš¶šŗš¶šš®šš¶š¼š» with a tic-tac-toe env
No critic model needed, the group is the baseline
Simpler than PPO
1ļøā£ Rollout generation: from the same board, model plays N games via sampling
2ļøā£ Each game scored with deterministic rewards (win, format, ...)
3ļøā£ Mean score computed across the group
4ļøā£ Each rollout's advantage = its score minus the group mean
5ļøā£ Model updated to favor trajectories above baseline
š Repeat
For a deep dive, check out
š± https://github.com/anakin87/llm-rl-environments-lil-course
a free hands-on course on RL environments for LLMs
It all starts with š„š²š¶š»š³š¼šæš°š²šŗš²š»š šš²š®šæš»š¶š»š“ šš¶ššµ š©š²šæš¶š³š¶š®šÆš¹š² š„š²šš®šæš±š
- question asked
- model generates reasoning + answer
- answer checked against ground truth
- reward drives RL training
In this setup, the environment is simple: fixed questions and answers, rollout logic, reward(s)
Consider a more complex tic-tac-toe env āā
It adds:
- dynamic game generation/handling
- tunable opponent skill
- multi-turn interactions
(envs can also include tools)
---
What happens at training?
We use ššæš¼šš½ š„š²š¹š®šš¶šš² š£š¼š¹š¶š°š š¢š½šš¶šŗš¶šš®šš¶š¼š» with a tic-tac-toe env
No critic model needed, the group is the baseline
Simpler than PPO
1ļøā£ Rollout generation: from the same board, model plays N games via sampling
2ļøā£ Each game scored with deterministic rewards (win, format, ...)
3ļøā£ Mean score computed across the group
4ļøā£ Each rollout's advantage = its score minus the group mean
5ļøā£ Model updated to favor trajectories above baseline
š Repeat
For a deep dive, check out
š± https://github.com/anakin87/llm-rl-environments-lil-course
a free hands-on course on RL environments for LLMs
- 2 replies
commented a
paper 8 months ago
liked a
Space 8 months ago
reacted to MonsterMMORPG's post with ā¤ļø 8 months ago
Post
4049
Compared Quality and Speed Difference (with CUDA 13 & Sage Attention) of BF16 vs GGUF Q8 vs FP8 Scaled vs NVFP4 for Z Image Turbo, FLUX Dev, FLUX SRPO, FLUX Kontext, FLUX 2 - Full 4K step by step tutorial also published
Full 4K tutorial : https://youtu.be/XDzspWgnzxI
Check above full 4K tutorial to learn more and see uncompressed original quality and size images
It was always wondered how much quality and speed difference exists between BF16, GGUF, FP8 Scaled and NVFP4 precisions. In this tutorial I have compared all these precision and quantization variants for both speed and quality. The results are pretty surprising. Moreover, we have developed and published NVFP4 model quant generator app and FP8 Scaled quant generator apps. The links of the apps are below if you want to use them. Furthermore, upgrading ComfyUI to CUDA 13 with properly compiled libraries is now very much recommended. We have observed some noticeable performance gains with CUDA 13. So for both SwarmUI and ComfyUI solo users, CUDA 13 ComfyUI is now recommended.
Full 4K tutorial : https://youtu.be/XDzspWgnzxI
Check above full 4K tutorial to learn more and see uncompressed original quality and size images
It was always wondered how much quality and speed difference exists between BF16, GGUF, FP8 Scaled and NVFP4 precisions. In this tutorial I have compared all these precision and quantization variants for both speed and quality. The results are pretty surprising. Moreover, we have developed and published NVFP4 model quant generator app and FP8 Scaled quant generator apps. The links of the apps are below if you want to use them. Furthermore, upgrading ComfyUI to CUDA 13 with properly compiled libraries is now very much recommended. We have observed some noticeable performance gains with CUDA 13. So for both SwarmUI and ComfyUI solo users, CUDA 13 ComfyUI is now recommended.
- 5 replies
liked a
Space 8 months ago
Where can I see the complete diagram?
reacted to AdinaY's post with š„ 8 months ago
Post
2617
Based on 2025 Chinese AI Timeline, here are some interesting takeaways:
āØ DeepSeek cadence: They shipped almost every month! (except Feb 2025)
āØ Qwen trajectory: Not a single āhitā model, but an expanding product line. VL/Math/Coder/Reranker/Embedding/Omni/Next/Image
āØ Multimodal trend: Steadily rising share, shifting from generation to editing + tooling.
āØ Reasoning as a main track: more engineered, system-level reasoning.
āØ From foundation to components: growth in infra models (embeddings, rerankers, OCR, speech) signals a move toward deployable stacks.
āØ Ecosystem broadening: more players beyond the top labs.
Follow for more updatesš
zh-ai-community
āØ DeepSeek cadence: They shipped almost every month! (except Feb 2025)
āØ Qwen trajectory: Not a single āhitā model, but an expanding product line. VL/Math/Coder/Reranker/Embedding/Omni/Next/Image
āØ Multimodal trend: Steadily rising share, shifting from generation to editing + tooling.
āØ Reasoning as a main track: more engineered, system-level reasoning.
āØ From foundation to components: growth in infra models (embeddings, rerankers, OCR, speech) signals a move toward deployable stacks.
āØ Ecosystem broadening: more players beyond the top labs.
Follow for more updatesš
- 2 replies
liked a
model 10 months ago
upvoted a paper about 1 year ago
liked a
model about 1 year ago
how to run the model on mobile device with qualcomm soc
2
#5 opened about 1 year ago
by eoe
liked a
model about 1 year ago
liked a
model over 1 year ago
upvoted a collection over 1 year ago
liked 2
models over 1 year ago
can you share your export code?
4
#1 opened almost 2 years ago
by eoe