Spaces:
Paused
Apply for a GPU community grant: Academic project
Here’s copy‑paste text you can use for the Description (and a quick note on project type).
Project type
Pick Academic project only if you’re affiliated with a school/lab and the work is primarily research. If this is mainly a public OSS benchmark + Space + notebooks, open‑source / open‑science community (if that option exists) is often a better semantic fit—but use whatever matches their dropdown literally.
Description (suggested)
Title line (first sentence):
Open‑science RL benchmark for verifiable SaaS support triage (OpenEnv hackathon) — public Space + reproducible GRPO training.
Body:
We are building Support Ops Env, an open‑source OpenEnv environment with a deterministic multi‑component grader (investigation, routing, reply quality, groundedness, submission + structured penalties) and a public Hugging Face Space that hosts the simulator for remote rollouts.
The goal is a reproducible RL evaluation stack for “first‑line support operator” workflows across realistic multi‑app tools (inbox/CRM/billing/access/policy/comms), including curriculum learning and tooling for baseline vs trained evaluation, reward‑hacking audits, and training plots—all intended to be shared openly for the community to extend.
Why we need GPU support: GRPO rollouts against a remote Space are VRAM‑heavy (multi‑generation rollouts + multi‑turn tool use). A T4‑class 16GB setup is frequently OOM‑limited, which blocks completing credible training runs and publishing before/after metrics that make the benchmark scientifically useful.
Requested help: Free or discounted HF Space GPU time (ideally 1× L4 24GB or better for short windows) so we can finish end‑to‑end GRPO runs, export artifacts (logs/plots/eval tables), and publish a stronger open‑science baseline with the existing public demo.
Public artifacts
Live Space (Docker): https://huggingface.co/spaces/raj23211/support-ops-env
Training path: TRL GRPO + optional Unsloth; notebooks + train.py in the linked model repo / submission materials.
License / openness: code and environment are intended for public reuse; grant visibility in the community tab is OK.