version string | seed int64 | dev_frac float64 | method string | exclusions dict | n_universe int64 | n_dev int64 | cells dict | dev_ids list |
|---|---|---|---|---|---|---|---|---|
v1 | 20,260,730 | 0.2 | joint-cell stratified sampling over (bucket_20B|bucket_120B); cells with n>=3 get at least 1 dev member | {
"shards": [
"shard120b_v2_05",
"shard_v2_05"
],
"broken_listtyped_ids": 845,
"note": "all 845 broken ids excluded (808 in shard-05, 37 in shard-07)"
} | 9,660 | 1,931 | {
"0|0": {
"n": 596,
"n_dev": 119
},
"0|1-3": {
"n": 131,
"n_dev": 26
},
"0|12-15": {
"n": 6,
"n_dev": 1
},
"0|4-7": {
"n": 58,
"n_dev": 12
},
"0|8-11": {
"n": 17,
"n_dev": 3
},
"0|NA": {
"n": 3,
"n_dev": 1
},
"1-3|0": {
"n": 261,
"n_dev"... | [
"apps-01615",
"apps-01669",
"apps-01681",
"apps-01686",
"apps-01693",
"apps-01698",
"apps-01706",
"apps-01707",
"apps-01713",
"apps-01718",
"apps-01725",
"apps-01739",
"apps-01749",
"apps-01774",
"apps-01780",
"apps-01787",
"apps-01804",
"apps-01822",
"apps-01869",
"apps-01873"... |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
round2-oss-matched — 第二轮 4 组实验数据（每组 10 节点，共 40）
代码：repo 分支 claude/round2-matched-compute（先 git fetch origin && git merge origin/claude/round2-matched-compute）。
目录：
exp0_20b/node00..04/pool.jsonl # 实验 0:shard-05 修复重跑(20B)
exp0_120b/node00..04/pool.jsonl # 实验 0:同上(120B)
exp1_20b/node00..09/{seeds,budgets,pool}.jsonl # 实验 1:20B token 对齐独立采样
exp2_120b/node00..09/{seeds,budgets,pool}.jsonl # 实验 2:120B 同上
exp3_120b/node00..09/{ck_nonsat/,nonsat_seeds,budgets,pool,cfg_loops.yaml} # 实验 3:自测试反馈 SE
loop_budget_v2_adaptive.json # 已锁定的逐题 loop 预算(勿改)
① 起模型（每节点，和 v2 跑法一样）：20B 用 bash handoff/serve_20b.sh;120B 用现有 TP2×4 脚本，LB :8000。
② 每节点一条命令（只改 RUN_DIR / K，绝对路径，nohup 后台）：
实验 0（20B/120B 各 K=00..04）:
cd <tts-sft>
RUN_DIR=<DATA>/exp0_20b/node0K RUN_NAME=exp0_20b_node0K SHARD_POOL=$RUN_DIR/pool.jsonl \
LOOPS=10 nohup bash handoff/run_shard_v2_fb.sh > $RUN_DIR/run.log 2>&1 &
# 120B:exp0_120b / exp0_120b_node0K / handoff/run_shard_v2_120b_fb.sh
实验 1（K=00..09）:
cd <tts-sft>
RUN_DIR=<DATA>/exp1_20b/node0K
nohup ../envs/vllm-gptoss/bin/python handoff/round2/run_token_matched_sampling.py \
--seeds $RUN_DIR/seeds.jsonl --budgets $RUN_DIR/budgets.jsonl \
--out $RUN_DIR/run --model gpt-oss-20b --base-url http://localhost:8000/v1 \
> $RUN_DIR/run.log 2>&1 &
实验 2：同实验 1，改 exp2_120b 和 --model gpt-oss-120b。
实验 3（K=00..09）:
cd <tts-sft>
RUN_DIR=<DATA>/exp3_120b/node0K RUN_NAME=exp3_selffb_node0K \
nohup bash handoff/round2/run_shard_v2_120b_selffb.sh > $RUN_DIR/run.log 2>&1 &
（记得像 run_shard_v2_fb.sh 里那样 unset 代理。）
断点续跑 = 原样重跑同一条命令，自动接着走。
③ 回传：实验 0/3 传整个 $RUN_DIR;实验 1/2 只传 $RUN_DIR/run/samples.jsonl（+run.log）。判分在我们这边做。
- Downloads last month
- 519