| license: mit | |
| # GLYPH crate templates + RL prompts | |
| Companion data for the [`glyph`](https://github.com/JayZenith/GLYPH) | |
| verifiers environment (Prime Intellect Environments Hub). Not meant to be | |
| used standalone — `glyph_env.load_environment()` downloads and caches this | |
| automatically on first use. | |
| ## Contents | |
| - `rl_prompts_pool_b_mixed_oversampled.jsonl` — 703 RL training prompts, each | |
| a `{"prompt": [...], "case_id": ..., "blueprint_root": "blueprints/<case>", | |
| "trace_prefix": ..., "expected_tool": ..., "expected_output": ...}` row. | |
| - `blueprints/<case_id>/` — the real Rust crate template for each prompt | |
| (`Cargo.toml` + `src/`), copied into a fresh sandbox per rollout by the | |
| environment before the agent's tool calls run against it. | |
| See [`JayZenith/Glyph-RLVR-Eval-Results`](https://huggingface.co/datasets/JayZenith/Glyph-RLVR-Eval-Results) | |
| for the held-out eval crates and raw per-rollout results instead — this | |
| dataset is training-side only. | |