Skip to content

Actions: Yatimai/reasonforge

Actions

All workflows

Actions

Loading...
Loading

Showing runs from all workflows
35 workflow runs
35 workflow runs

Filter by Workflow

⌥ + click/return to exclude

Filter by Event

⌥ + click/return to exclude

Filter by Status

⌥ + click/return to exclude

Filter by Branch

⌥ + click/return to exclude

Filter by Actor

⌥ + click/return to exclude
ci: add minimal GitHub Actions workflow (test only) CI #1: Commit 8dd09a9 pushed by Yatimai
14s main
main
14s
fix: cache old_log_probs for real PPO signal + tune hyperparams Tests #34: Commit d84c7f1 pushed by Yatimai
27s main
main
27s
fix: GPU memory leak — model.cpu() before del to free VRAM Tests #33: Commit 6b84996 pushed by Yatimai
18s main
main
18s
feat: GRPO v2 — 7-module LoRA (43M params), re-mining, lower LR Tests #32: Commit 3117b32 pushed by Yatimai
14s main
main
14s
fix: per-sample backward + deferred model load for OOM-free GRPO Tests #31: Commit c15c8c7 pushed by Yatimai
20s main
main
20s
config: eval_every=50, save_every=100, mine full 7000 questions Tests #30: Commit c158631 pushed by Yatimai
19s main
main
19s
feat: GRPO hard example mining + OOM fixes + single model Tests #29: Commit 48897b2 pushed by Yatimai
17s main
main
17s
fix: use Schulman's unbiased KL estimator (always >= 0) Tests #28: Commit 3dcb115 pushed by Yatimai
16s main
main
16s
fix: sft_train.py eval + cleanup Tests #27: Commit ac246e0 pushed by Yatimai
23s main
main
23s
fix: GRPO robustness — device_map, vLLM validation, monitoring metrics Tests #26: Commit 68e5617 pushed by Yatimai
17s main
main
17s
feat: add vLLM batch generation for GRPO (Option B) Tests #25: Commit 9fdc195 pushed by Yatimai
20s main
main
20s
GRPO: subsample eval to 200 questions for faster training Tests #24: Commit f8ccb74 pushed by Yatimai
20s main
main
20s
GRPO: batch log_probs for ~5x speedup, reduce to 500 steps Tests #23: Commit 6f99133 pushed by Yatimai
18s main
main
18s
Align GRPO LoRA defaults with STaR/SFT config Tests #22: Commit ec6b9b7 pushed by Yatimai
21s main
main
21s
Fix fallback values and misleading comment Tests #21: Commit fcd225e pushed by Yatimai
20s main
main
20s
Tune STaR hyperparameters: canonical SQL selection, LoRA 16, 2 epochs Tests #20: Commit 8083ac0 pushed by Yatimai
22s main
main
22s
GRPO: batched generation, binary reward, full dev eval Tests #19: Commit e43c65f pushed by Yatimai
22s main
main
22s
Add eval_steps and switch to non-cumulative SFT data Tests #18: Commit 4d5327c pushed by Yatimai
20s main
main
20s
Update LoRA rank to 32 for STaR and GRPO Tests #17: Commit d74815a pushed by Yatimai
22s main
main
22s
Fix eval to only count exact match (reward >= 1.0) Tests #16: Commit e5ca4f2 pushed by Yatimai
20s main
main
20s
Fix CLI default --kl_coef to match config (0.01 → 0.001) Tests #15: Commit 34667b8 pushed by Yatimai
18s main
main
18s
Tune GRPO hyperparameters for H200 run Tests #14: Commit d7a3abb pushed by Yatimai
15s main
main
15s
Upgrade GRPO to v3 + tune hyperparameters for H200 run Tests #13: Commit db7bca8 pushed by Yatimai
16s main
main
16s
Fix bugs from code reviews #3 and #4 Tests #12: Commit b5386ec pushed by Yatimai
21s main
main
21s
Improve robustness and code quality Tests #11: Commit 90719c7 pushed by Yatimai
19s main
main
19s