Skip to content

Actions: AwareLiquid/C1

Actions

Experiments

Actions

Loading...
Loading

Show workflow options

Create status badge

Loading
experiments.yml
29 workflow runs
29 workflow runs

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
results(e9): DEAD — 去噪监督无外推增量(三模式外推全随机, 加噪变体域内更差) Experiments #29: Commit a15d422 pushed by everest-an
1m 34s main
main
1m 34s
fix(e9): min_len 8->1 — e3 从 k=1 起训练; v4 跳过短 parity 起手难不进 grok Experiments #28: Commit 8af976b pushed by everest-an
1m 13s main
main
1m 13s
fix(e9): v4 gen_parity 端到端 — v3 手写布局与模型答案槽语义错位 (loss 卡 log2 地板) Experiments #27: Commit bebf8c5 pushed by everest-an
1m 22s main
main
1m 22s
fix(e9): 答案-only 监督 — v1/v2 全位置 CE 稀释答案梯度 1/32 不进 grok (ADJ-001 budge… Experiments #26: Commit 3d61a79 pushed by everest-an
1m 32s main
main
1m 32s
results(s11): E5 h=512 全视野统一头优势 — h8/32/128/512 均优于专用(-4.3%~-9.3%) Experiments #25: Commit 0788d9c pushed by everest-an
1m 46s main
main
1m 46s
feat(e9)+docs: 去噪监督试点 + 组件立项 — 换掉 next-token 的范式探针 Experiments #24: Commit 315845f pushed by everest-an
1m 25s main
main
1m 25s
feat(e9)+docs: 去噪监督试点 + 组件立项 — 换掉 next-token 的范式探针 Experiments #23: Commit 315845f pushed by everest-an
1m 35s main
main
1m 35s
results(e7): 探针 A DEAD — 事件边界形态对语言线无受力面(boundary +4.1 PPL, 2/3 配对更差, … Experiments #22: Commit 160266a pushed by everest-an
1m 9s main
main
1m 9s
fix(e7): token 生成越界 — choice(vocab_size)+1 产生 token=vocab_size 越界, GP… Experiments #21: Commit 96268eb pushed by everest-an
1m 17s main
main
1m 17s
results(s10): E5 @1040/16L 长预算 8000 步 — 统一头优势恢复, S8 反转=欠训练 Experiments #20: Commit ae53ca5 pushed by everest-an
1m 16s main
main
1m 16s
docs(e7,e8)+feat: 数据源=L1 定性 — 数据形态探针 + 数据生成闭环设计 Experiments #19: Commit f63f197 pushed by everest-an
1m 21s main
main
1m 21s
docs: C1 实验总览(S1-S10 汇总 + 5 个 bug 清单 + 方法学备注) Experiments #18: Commit dee6f60 pushed by everest-an
1m 8s main
main
1m 8s
results: S9 结果入库 — E5 长视野优势跨预算稳健(短视野 h=8 为噪声区) Experiments #17: Commit 37c941a pushed by everest-an
1m 13s main
main
1m 13s
results: S8 结果入库 — E5 模型放大 1040/16L 后统一头优势未保持(反转) Experiments #16: Commit 2a53f92 pushed by everest-an
1m 20s main
main
1m 20s
results: S7b 结果入库 — E5 dedicated 重跑 @4000(统一头全视野更优)+ 长视野 512 Experiments #15: Commit 16adbd7 pushed by everest-an
1m 17s main
main
1m 17s
results: S7 结果入库 — E5 视野边界 bug + 权重模式消融(inv_sqrt 最优) Experiments #14: Commit 50a42ed pushed by everest-an
1m 8s main
main
1m 8s
fix(e5): 修复视野 >= seq_len 时的退化(loss 变 float + dedicated 崩溃) Experiments #13: Commit 09ef3ff pushed by everest-an
1m 10s main
main
1m 10s
results: S6 结果入库 — E5 统一主干优势随视野增长(h=16 峰值 15.5%) Experiments #12: Commit bbdb4d7 pushed by everest-an
1m 4s main
main
1m 4s
results: S5 结果入库 — E5 6-seed 确证(统一主干利于长视野) / E6 5-seed 确证(域依赖零样本) Experiments #11: Commit 90cbe7f pushed by everest-an
1m 8s main
main
1m 8s
results: S4 结果入库 — E5 统一主干利于长视野 / E6 域依赖零样本迁移 Experiments #10: Commit fbcbfe5 pushed by everest-an
1m 14s main
main
1m 14s
results: S3 结果入库 — E1 GRPO 稳定化修复成功 / E3 尺度信号仍弱 Experiments #9: Commit 2e69a6c pushed by everest-an
1m 12s main
main
1m 12s
fix(e1): GRPO 优势稳定化(std 下限 + 裁剪 + KL 锚定) Experiments #8: Commit 8e34960 pushed by everest-an
1m 3s main
main
1m 3s
fix(e3): 复用 M1 的 gen_parity 协议 + 加训练 loss 打印 Experiments #7: Commit d44ad73 pushed by everest-an
1m 15s main
main
1m 15s
feat(e3): 加 --selective 开关(默认开)以套用 M1 已验证 parity 配方 Experiments #6: Commit 9682cb6 pushed by everest-an
1m 8s main
main
1m 8s
fix(e2): MTP 头改为预测未来 token(t+j+1) + 评估用已训练 head[0] Experiments #5: Commit 0520ce0 pushed by everest-an
1m 7s main
main
1m 7s