Add EmailTriageEnv: Real-World OpenEnv for Multi-Task Email Triage#499

Open

Rhushya wants to merge 60 commits into

huggingface:mainfrom

Commits on Apr 4, 2026

Commits on Apr 24, 2026

Round 2: Oversight Inbox Arena - Multi-agent multi-turn environment
Rhushya
committed

Commits on Apr 25, 2026

feat: Add interactive Gradio UI for Hugging Face Space
Rhushya
committed
fix: remove root route to allow Gradio UI to mount
Rhushya
committed
fix: unpack error in gradio callbacks
Rhushya
committed
Update environment and UI
Rhushya
committed
fix: dynamically check bf16 support for older GPUs like Colab T4
Rhushya
committed
fix: remove environment_factory from GRPOTrainer and parse rewards dynamically via XML
Rhushya
committed
fix: handle list-type completions in reward functions
Rhushya
committed
fix: load model in 4-bit via Unsloth to prevent Colab T4 OOM
Rhushya
committed
fix: extreme VRAM optimizations for Colab T4 (8-bit paged adamw, smaller lora, smaller seq length)
Rhushya
committed
fix: ultra low VRAM profile to guarantee Colab T4 compatibility
Rhushya
committed
fix: switch to Qwen2-0.5B, complete rewrite for guaranteed Colab T4 compatibility
Rhushya
committed
feat: premium Gradio UI on HF Space + push-to-hub training support
Rhushya
committed
fix(space): pin gradio below v6 for ui compatibility
Rhushya
committed
fix(ui): remove unsupported markdown scale args for gradio
Rhushya
committed
Clean email triage PR artifacts and stabilize demo workflow.
Rhushya
committed
Fix email triage runtime issues and test integration.
Rhushya
committed
Finalize showcase docs and Hugging Face Space template.
Rhushya
committed
Stabilize Colab T4 GRPO workflow and notebook.
Rhushya
committed
Fix Colab T4 import failures and finalize training notebook.
Rhushya
committed
fix: remove max_prompt_length from GRPOConfig (dropped in TRL v1.0+)
Rhushya
committed
chore: add Colab training notebook for email triage GRPO
Rhushya
committed
fix: unblock GRPO learning — longer completions, reward variance, real email prompts
Rhushya
committed
feat: switch default model to Qwen2.5-1.5B, bump max_completion_length to 256
Rhushya
committed
fix: remove JSON BOM, add temperature=0.9, bump completion/prompt lengths for Qwen2.5-1.5B
Rhushya
committed
chore: update training notebooks — latest Colab version with all fixes
Rhushya
committed
fix: robust Environment import fallback — no longer requires fastmcp for training
Rhushya
committed
fix: auto-install mergekit and fastmcp before TRL import
Rhushya
committed
fix: proper reward extraction in easy mode + robust import fallbacks for models.py
Rhushya
committed
chore: clean production notebook — pinned deps, no smoke-test conflicts, all cells tested
Rhushya
committed
Add simple GRPO training notebook (no vLLM, Unsloth 4-bit, T4-safe)
Rhushya
committed
Replace UI with clean black-and-white Gradio theme
Rhushya
committed
style: force white background on ALL Gradio elements — no dark panels
Rhushya
committed
Enforce black-on-white Gradio UI styling
Rhushya
committed

Commits on Apr 26, 2026