Hugging Face's logo

tchbcb
/
samai-2b

Safetensors
llama
custom_code
Model card Files
xet
 Community
samai-2b
File size: 3,621 Bytes
504da3d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
{
  "contract_version": "r5b-2-integrated",
  "declared_at": "2026-09-09",
  "supersedes": "r5b-1",
  "declared_repo_id": "tchbcn/samai-pnet-dmoe-2b",
  "actual_repo_id": "tchbcb/samai-pnet-dmoe-2b",
  "note": "用户指定仓名 tchbcn/samai-pnet-dmoe-2b; 所持 token 实际属主为 tchbcb (whoami 实证, tchbcn 命名空间 403 无权创建), 故以 tchbcb/ 落库, 仓内容与命名意图一致",
  "release": {
    "model_file": "model.safetensors (全量合并: 基座+LoRA+halting头+Router/MTP)",
    "dtype_policy": "主干/专家 fp16; ponder_head fp32 (round4 教训)",
    "trust_remote_code": true,
    "requirements": "torch>=2.4, transformers>=5.0, peft 仅训练侧需要"
  },
  "lineage": {
    "base": "tchbcb/MiniCPM5-2B-cpu (42L/H2048/I6144/GQA16+2/V130560)",
    "convert": "pnet2b_moe: 思考块[34,42) 8层x8同构专家, 参数守恒",
    "round7": "ckpt_2b (宪法四损失基线)",
    "round5B": "ckpt_2b_r5b: per-step CE 0.5 + difficulty prior, init-from ckpt_2b"
  },
  "frozen_structure": {
    "ponder_start_layer": 34,
    "max_ponder_steps": 8,
    "ponder_window": 1,
    "moe_num_experts": 8,
    "moe_scope": "ponder",
    "mtp_num_tokens": 2
  },
  "inference_semantics": {
    "halting": "learned head (fp32, sigmoid), bias 已训练",
    "stop_rule": "cum(w) >= 1 - eps 即停, 残余质量记给最后执行步",
    "epsilon": 0.01,
    "tau": 0.125,
    "route_rule": "每 token 仅第 1 轮迭代 fresh 路由一次, 循环/commit 全程 frozen 复用",
    "moe_activation": "p_i >= tau 激活, 激活数 1~8 自适应, 无一过阈 top-1 保底",
    "kv": "commit_kv=true: 混合表示经 commit pass 写最终 KV",
    "mtp": "训练期增强, 推理 generate() 自动旁路"
  },
  "eval_anchors_90rows_zh_cot": {
    "round7_baseline": {
      "hard": {
        "ce": 4.073070367177327,
        "active": 1.6183046340942382,
        "steps": 2.066666666666667,
        "mtp": 0.5716088751951853,
        "n": 15
      },
      "easy": {
        "ce": 4.018902349472046,
        "active": 1.6024614969889324,
        "steps": 2.0,
        "mtp": 0.5493664860725402,
        "n": 15
      },
      "medium": {
        "ce": 4.228601948420207,
        "active": 1.614480972290039,
        "steps": 2.033333333333333,
        "mtp": 0.5261600712935129,
        "n": 15
      }
    },
    "round5B_live_reference": {
      "hard": {
        "ce": 3.3185231844584147,
        "active": 1.5587466398874918,
        "steps": 2.8666666666666667,
        "mtp": 0.6124512155850729,
        "n": 15
      },
      "easy": {
        "ce": 3.264301141103109,
        "active": 1.5368041038513183,
        "steps": 2.433333333333333,
        "mtp": 0.6108810146649678,
        "n": 15
      },
      "medium": {
        "ce": 3.5550480524698895,
        "active": 1.5593219121297202,
        "steps": 2.6,
        "mtp": 0.5710210422674815,
        "n": 15
      }
    },
    "round5B_merged_release": {
      "hard": {
        "ce": 3.3181373119354247,
        "steps": 2.8666666666666667,
        "n": 15
      },
      "easy": {
        "ce": 3.2644861777623495,
        "steps": 2.433333333333333,
        "n": 15
      },
      "medium": {
        "ce": 3.555168342590332,
        "steps": 2.6,
        "n": 15
      }
    },
    "eval_protocol": "eval_ponder_moe.py 口径: use_cache=False, batch_size=2, fp16, 90 条 zh_cot",
    "merge_note": "干净进程终验 (verify_release.json, trust_remote_code): ΔCE≤0.001, steps 三桶与 live 参考一致 (2.433/2.600/2.867), 分化 +0.433 完整保留; 同进程对分探针曾见 ≤0.065 CE 交叉污染, 已排除"
  }
}