Spaces:
Paused
Paused
Commit History
-
Fix torch bfloat16 errors on T4 GPUs by enabling Unsloth dtype auto-detection and explicitly wrapping forward passes in autocast
4b22b06 -
Add mandatory Unsloth inference state toggles around generation for RL pipeline
81ed883 -
Capture and display exact Unsloth import exception
fbf8187 -
Update with existing environment
d81b76a -
Fix GRPO group-loss training and align UI defaults.
Train on all K samples per group for GRPO updates, and update Gradio defaults to group_size 8 with max generation tokens 60 to match the new training configuration.5a0c6af -
Migrate LLM pipeline to custom GRPO with robust rewards
Replace the REINFORCE-style loop with grouped GRPO optimization and add decomposed reward telemetry, GRPO diagnostics, and UI controls so training is more stable, observable, and reproducible.dfc5996
-
Multi-model benchmark pipeline: VRAM cleanup + EMA graph + detailed output
af6bbef -
Fix truncation: 80 tokens, regex safety net, strict prompt
deef82c -
Hackathon speedrun: max_new_tokens=32, seq_len=512 for 4-8x faster iterations
ee5ddee -
Replace flash-attn with PyTorch built-in SDPA (no CUDA compile needed)
e9dea07 -
Max GPU utilization: flash-attn2 + grad accumulation + 15 steps/ep + 1024 seq len
93d0171 -
Switch to Llama 3.1 8B + fix low-timestep crash (min 5000)
8d95050 -
Add LLM RL training with Gemma 7B + LoRA
ee3dfa7 -
Add Cloud Arena Mathematical Model RL environment
12263fa
