5.31 GB

Ctrl+K

1 contributor

History: 60 commits

tchbcb

round5-B 2B: per-step CE (hard-easy steps +0.067 -> +0.433, CE -0.8) + serve_contract r5b-1

d44095e 10 minutes ago

configs
PonderNet x MiniCPM5-2B: modeling + train/eval/probe/RL, configs for round4-5, docs (round4 +0.498 hard-easy) about 20 hours ago
data
round5: OOD 困难集 about 11 hours ago
data_sample
PonderNet x MiniCPM5-2B: modeling + train/eval/probe/RL, configs for round4-5, docs (round4 +0.498 hard-easy) about 20 hours ago
dev
ping test 43 minutes ago
docs
round5-B 2B: per-step CE (hard-easy steps +0.067 -> +0.433, CE -0.8) + serve_contract r5b-1 10 minutes ago
pondernet
round5-B 2B: per-step CE (hard-easy steps +0.067 -> +0.433, CE -0.8) + serve_contract r5b-1 10 minutes ago
.gitattributes

140 Bytes
round7 2B: weights via git-lfs (model 5.08GB + ckpt safetensors) 38 minutes ago
.gitignore

199 Bytes
round7 2B: 2B weights (model 5.08GB fp16 + Router/MTP/Halt/LoRA ckpt) 30 minutes ago
ARCHITECTURE.md

6.85 kB
round5: 架构宪法 (官方系统提示词) about 11 hours ago
README.md

29.7 kB
round5: 项目说明与实验报告 about 11 hours ago
eval_ponder.py

11.4 kB
round5: 评测: 分桶CE/步数分布 + forced-sweep 深度-质量曲线 about 11 hours ago
make_gsm8k_zh_data.py

4.79 kB
round5: GSM8K-zh 数据生成 about 11 hours ago
ponder_llama.py

27.3 kB
round5: 核心: PonderLlamaForCausalLM (循环思考+自适应停机+三方向开关) about 11 hours ago
ponder_steps_demo.py

2.65 kB
round5: 训练轨迹演示 about 11 hours ago
requirements.txt

75 Bytes
PonderNet x MiniCPM5-2B: modeling + train/eval/probe/RL, configs for round4-5, docs (round4 +0.498 hard-easy) about 20 hours ago
run_round5.sh

3.93 kB
round5: 第五轮实验总控脚本 about 11 hours ago
train_head_cached.py

16.9 kB
round5: 缓存训练: 特征缓存 head-only + 消融①软硬目标 about 11 hours ago
train_ponder_head.py

42.9 kB
round5: 全前向训练: head / head+lora / head+block + 方向①②开关 about 11 hours ago
train_rl_ponder.py

18 kB
round5: 方向③: 答案级 RL (可验证奖励 head-only REINFORCE) about 11 hours ago