| tags: | |
| - rlvr | |
| - grpo | |
| - rust | |
| - tool-use | |
| - passk | |
| # RLVR_HELDOUT69_PASSK_STEP25_DATASET | |
| This is the exact 8-prompt RLVR training dataset used for `JayZenith/RLVR_HELDOUT69_PASSK_STEP25`. | |
| Source: original held-out-69 failures rescanned with SFT_V1 pass@8. Kept only prompts where SFT_V1 showed latent capability under sampling (`0 < solves < 8`). | |
| Files: | |
| ```text | |
| rl_prompts_heldout69_passk_targets.jsonl # 8 PRIME-RL/verifiers training rows | |
| heldout69_passk_target_names.txt # prompt/case ids used for matched pass@8 eval | |
| ``` | |
| Matched evaluation on the resulting step-25 checkpoint: | |
| ```text | |
| SFT_V1: 47/64 = 0.734 | |
| step_25: 54/64 = 0.844 | |
| delta: +7 solves, +10.9 pts | |
| ``` | |
| Training command and eval command are documented in the glyph repo README and `blog/blog_copy_copy.md`. | |
| Scope note: this dataset is intentionally narrow. It is for testing whether RLVR improves reliability on prompts the SFT policy can already sometimes solve, not for broad Rust-agent generalization. | |