Orion checkpoint phase5-step-004900
Browse files- checkpoints/phase5-step-004900/config.json +45 -0
- checkpoints/phase5-step-004900/manifest.json +405 -0
- checkpoints/phase5-step-004900/model-0000.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0001.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0002.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0003.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0004.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0005.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0006.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0007.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0008.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0009.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0010.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0011.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0012.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0013.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0014.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0015.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0016.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0017.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0018.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0019.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0020.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0021.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0022.safetensors +3 -0
- checkpoints/phase5-step-004900/model-0023.safetensors +3 -0
- checkpoints/phase5-step-004900/orion_phase5_launcher.py +894 -0
- checkpoints/phase5-step-004900/orion_unified_trainer.py +0 -0
- checkpoints/phase5-step-004900/tokenizer/tokenizer.json +0 -0
- checkpoints/phase5-step-004900/tokenizer/tokenizer_config.json +17 -0
checkpoints/phase5-step-004900/config.json
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_name": "Orion Flagship 2B T2.2",
|
| 3 |
+
"implementation_version": 5,
|
| 4 |
+
"architecture": "orion_flagship_2b_t2_2",
|
| 5 |
+
"d_model": 2112,
|
| 6 |
+
"n_layers": 24,
|
| 7 |
+
"n_heads": 24,
|
| 8 |
+
"n_kv_heads": 6,
|
| 9 |
+
"rope_theta": 10000.0,
|
| 10 |
+
"sequence_length": 32768,
|
| 11 |
+
"tokenizer_name": "mistralai/Mistral-7B-v0.3",
|
| 12 |
+
"n_procedure_banks": 6,
|
| 13 |
+
"procedure_bank_span": 4,
|
| 14 |
+
"n_experts": 12,
|
| 15 |
+
"top_k": 1,
|
| 16 |
+
"expert_hidden": 3456,
|
| 17 |
+
"shared_expert_hidden": 1280,
|
| 18 |
+
"vault_key_parts": 128,
|
| 19 |
+
"vault_slots": 16384,
|
| 20 |
+
"vault_top_component": 12,
|
| 21 |
+
"vault_top_k": 4,
|
| 22 |
+
"vault_gate_groups": 64,
|
| 23 |
+
"vault_read_layers": [
|
| 24 |
+
3,
|
| 25 |
+
7,
|
| 26 |
+
11,
|
| 27 |
+
15,
|
| 28 |
+
19,
|
| 29 |
+
23
|
| 30 |
+
],
|
| 31 |
+
"working_state_layers": [
|
| 32 |
+
3,
|
| 33 |
+
7,
|
| 34 |
+
11,
|
| 35 |
+
15,
|
| 36 |
+
19,
|
| 37 |
+
23
|
| 38 |
+
],
|
| 39 |
+
"deliberation_start": 18,
|
| 40 |
+
"deliberation_cycles": 1,
|
| 41 |
+
"vocab_size": 32768,
|
| 42 |
+
"rope_extension": "yarn_frequency_blend_v1",
|
| 43 |
+
"rope_8k_frequency_preserved": true,
|
| 44 |
+
"phase4_max_train_context": 32768
|
| 45 |
+
}
|
checkpoints/phase5-step-004900/manifest.json
ADDED
|
@@ -0,0 +1,405 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"format_version": "orion-phase5-1",
|
| 3 |
+
"checkpoint_format": "single_gpu_sft",
|
| 4 |
+
"model_shards": [
|
| 5 |
+
"model-0000.safetensors",
|
| 6 |
+
"model-0001.safetensors",
|
| 7 |
+
"model-0002.safetensors",
|
| 8 |
+
"model-0003.safetensors",
|
| 9 |
+
"model-0004.safetensors",
|
| 10 |
+
"model-0005.safetensors",
|
| 11 |
+
"model-0006.safetensors",
|
| 12 |
+
"model-0007.safetensors",
|
| 13 |
+
"model-0008.safetensors",
|
| 14 |
+
"model-0009.safetensors",
|
| 15 |
+
"model-0010.safetensors",
|
| 16 |
+
"model-0011.safetensors",
|
| 17 |
+
"model-0012.safetensors",
|
| 18 |
+
"model-0013.safetensors",
|
| 19 |
+
"model-0014.safetensors",
|
| 20 |
+
"model-0015.safetensors",
|
| 21 |
+
"model-0016.safetensors",
|
| 22 |
+
"model-0017.safetensors",
|
| 23 |
+
"model-0018.safetensors",
|
| 24 |
+
"model-0019.safetensors",
|
| 25 |
+
"model-0020.safetensors",
|
| 26 |
+
"model-0021.safetensors",
|
| 27 |
+
"model-0022.safetensors",
|
| 28 |
+
"model-0023.safetensors"
|
| 29 |
+
],
|
| 30 |
+
"step": 4900,
|
| 31 |
+
"architecture": {
|
| 32 |
+
"model_name": "Orion Flagship 2B T2.2",
|
| 33 |
+
"implementation_version": 5,
|
| 34 |
+
"architecture": "orion_flagship_2b_t2_2",
|
| 35 |
+
"d_model": 2112,
|
| 36 |
+
"n_layers": 24,
|
| 37 |
+
"n_heads": 24,
|
| 38 |
+
"n_kv_heads": 6,
|
| 39 |
+
"rope_theta": 10000.0,
|
| 40 |
+
"sequence_length": 32768,
|
| 41 |
+
"tokenizer_name": "mistralai/Mistral-7B-v0.3",
|
| 42 |
+
"n_procedure_banks": 6,
|
| 43 |
+
"procedure_bank_span": 4,
|
| 44 |
+
"n_experts": 12,
|
| 45 |
+
"top_k": 1,
|
| 46 |
+
"expert_hidden": 3456,
|
| 47 |
+
"shared_expert_hidden": 1280,
|
| 48 |
+
"vault_key_parts": 128,
|
| 49 |
+
"vault_slots": 16384,
|
| 50 |
+
"vault_top_component": 12,
|
| 51 |
+
"vault_top_k": 4,
|
| 52 |
+
"vault_gate_groups": 64,
|
| 53 |
+
"vault_read_layers": [
|
| 54 |
+
3,
|
| 55 |
+
7,
|
| 56 |
+
11,
|
| 57 |
+
15,
|
| 58 |
+
19,
|
| 59 |
+
23
|
| 60 |
+
],
|
| 61 |
+
"working_state_layers": [
|
| 62 |
+
3,
|
| 63 |
+
7,
|
| 64 |
+
11,
|
| 65 |
+
15,
|
| 66 |
+
19,
|
| 67 |
+
23
|
| 68 |
+
],
|
| 69 |
+
"deliberation_start": 18,
|
| 70 |
+
"deliberation_cycles": 1,
|
| 71 |
+
"vocab_size": 32768,
|
| 72 |
+
"rope_extension": "yarn_frequency_blend_v1",
|
| 73 |
+
"rope_8k_frequency_preserved": true,
|
| 74 |
+
"phase4_max_train_context": 32768
|
| 75 |
+
},
|
| 76 |
+
"sources": {
|
| 77 |
+
"minicpm_agent_code": {
|
| 78 |
+
"repo": "minicpm_agent_code",
|
| 79 |
+
"raw_rows": 7606
|
| 80 |
+
},
|
| 81 |
+
"minicpm_agent_general": {
|
| 82 |
+
"repo": "minicpm_agent_general",
|
| 83 |
+
"raw_rows": 33957
|
| 84 |
+
},
|
| 85 |
+
"minicpm_agent_search": {
|
| 86 |
+
"repo": "minicpm_agent_search",
|
| 87 |
+
"raw_rows": 722
|
| 88 |
+
},
|
| 89 |
+
"hermes_multi": {
|
| 90 |
+
"repo": "NousResearch/hermes-function-calling-v1",
|
| 91 |
+
"raw_rows": 372
|
| 92 |
+
},
|
| 93 |
+
"hermes_single": {
|
| 94 |
+
"repo": "hermes_single",
|
| 95 |
+
"raw_rows": 318
|
| 96 |
+
},
|
| 97 |
+
"openai_schema_multi": {
|
| 98 |
+
"repo": "openai_schema_multi",
|
| 99 |
+
"raw_rows": 331
|
| 100 |
+
},
|
| 101 |
+
"smoltalk": {
|
| 102 |
+
"repo": "smoltalk",
|
| 103 |
+
"raw_rows": 565
|
| 104 |
+
},
|
| 105 |
+
"openr1_verified": {
|
| 106 |
+
"repo": "openr1_verified",
|
| 107 |
+
"raw_rows": 3023
|
| 108 |
+
},
|
| 109 |
+
"glm_reasoning": {
|
| 110 |
+
"repo": "glm_reasoning",
|
| 111 |
+
"raw_rows": 3397
|
| 112 |
+
},
|
| 113 |
+
"gptoss_reasoning": {
|
| 114 |
+
"repo": "gptoss_reasoning",
|
| 115 |
+
"raw_rows": 945
|
| 116 |
+
},
|
| 117 |
+
"claude_opus46_inversion": {
|
| 118 |
+
"repo": "claude_opus46_inversion",
|
| 119 |
+
"raw_rows": 1587
|
| 120 |
+
},
|
| 121 |
+
"claude_opus47_inversion": {
|
| 122 |
+
"repo": "claude_opus47_inversion",
|
| 123 |
+
"raw_rows": 1593
|
| 124 |
+
},
|
| 125 |
+
"claude_opus48_max": {
|
| 126 |
+
"repo": "claude_opus48_max",
|
| 127 |
+
"raw_rows": 285600
|
| 128 |
+
},
|
| 129 |
+
"high_coder": {
|
| 130 |
+
"repo": "high_coder",
|
| 131 |
+
"raw_rows": 478446
|
| 132 |
+
}
|
| 133 |
+
},
|
| 134 |
+
"optimizer_saved": false,
|
| 135 |
+
"stop_format": "User: {prompt}\nAssistant: {completion}<tokenizer.eos_token_id>",
|
| 136 |
+
"training_stage": "phase5",
|
| 137 |
+
"phase5_fingerprint": {
|
| 138 |
+
"version": "orion-phase5-minicpm-agent-v1",
|
| 139 |
+
"architecture": "orion_flagship_2b_t2_2",
|
| 140 |
+
"identity": "Orion T2.1 2B",
|
| 141 |
+
"phase4_parent": "checkpoints/phase4-step-000900",
|
| 142 |
+
"steps": 5000,
|
| 143 |
+
"lr": 8e-07,
|
| 144 |
+
"grad_accum": 4,
|
| 145 |
+
"max_context": 32768,
|
| 146 |
+
"sources": {
|
| 147 |
+
"minicpm_core": {
|
| 148 |
+
"repo": "openbmb/UltraData-SFT-2605",
|
| 149 |
+
"config": null,
|
| 150 |
+
"kind": "core",
|
| 151 |
+
"rights": "public",
|
| 152 |
+
"weight": 1.0
|
| 153 |
+
},
|
| 154 |
+
"minicpm_agent_code": {
|
| 155 |
+
"repo": "openbmb/UltraData-SFT-Agent-2609",
|
| 156 |
+
"config": "Code-Agent",
|
| 157 |
+
"kind": "agent",
|
| 158 |
+
"rights": "public",
|
| 159 |
+
"weight": 1.0
|
| 160 |
+
},
|
| 161 |
+
"minicpm_agent_general": {
|
| 162 |
+
"repo": "openbmb/UltraData-SFT-Agent-2609",
|
| 163 |
+
"config": "General-Agent",
|
| 164 |
+
"kind": "agent",
|
| 165 |
+
"rights": "public",
|
| 166 |
+
"weight": 1.2
|
| 167 |
+
},
|
| 168 |
+
"minicpm_agent_search": {
|
| 169 |
+
"repo": "openbmb/UltraData-SFT-Agent-2609",
|
| 170 |
+
"config": "Search-Agent",
|
| 171 |
+
"kind": "agent",
|
| 172 |
+
"rights": "public",
|
| 173 |
+
"weight": 0.8
|
| 174 |
+
},
|
| 175 |
+
"hermes_multi": {
|
| 176 |
+
"repo": "NousResearch/hermes-function-calling-v1",
|
| 177 |
+
"config": "func_calling",
|
| 178 |
+
"kind": "agent",
|
| 179 |
+
"rights": "public",
|
| 180 |
+
"weight": 0.6
|
| 181 |
+
},
|
| 182 |
+
"hermes_single": {
|
| 183 |
+
"repo": "NousResearch/hermes-function-calling-v1",
|
| 184 |
+
"config": "func_calling_singleturn",
|
| 185 |
+
"kind": "agent",
|
| 186 |
+
"rights": "public",
|
| 187 |
+
"weight": 0.35
|
| 188 |
+
},
|
| 189 |
+
"openai_schema_multi": {
|
| 190 |
+
"repo": "stindardlogic/openai-function-calling-5k",
|
| 191 |
+
"config": null,
|
| 192 |
+
"kind": "agent",
|
| 193 |
+
"rights": "public",
|
| 194 |
+
"weight": 0.45
|
| 195 |
+
},
|
| 196 |
+
"smoltalk": {
|
| 197 |
+
"repo": "HuggingFaceTB/smoltalk",
|
| 198 |
+
"config": "everyday-conversations",
|
| 199 |
+
"kind": "other",
|
| 200 |
+
"rights": "public",
|
| 201 |
+
"weight": 1.0
|
| 202 |
+
},
|
| 203 |
+
"openr1_verified": {
|
| 204 |
+
"repo": "open-r1/OpenR1-Math-220k",
|
| 205 |
+
"config": "default",
|
| 206 |
+
"kind": "other",
|
| 207 |
+
"rights": "public",
|
| 208 |
+
"weight": 0.8
|
| 209 |
+
},
|
| 210 |
+
"glm_reasoning": {
|
| 211 |
+
"repo": "Jackrong/GLM-5.1-Reasoning-1M-Cleaned",
|
| 212 |
+
"config": "main",
|
| 213 |
+
"kind": "other",
|
| 214 |
+
"rights": "other",
|
| 215 |
+
"weight": 1.5
|
| 216 |
+
},
|
| 217 |
+
"gptoss_reasoning": {
|
| 218 |
+
"repo": "Jackrong/ShareGPT-gpt-oss-120B-reasoning",
|
| 219 |
+
"config": null,
|
| 220 |
+
"kind": "other",
|
| 221 |
+
"rights": "other",
|
| 222 |
+
"weight": 1.0
|
| 223 |
+
},
|
| 224 |
+
"claude_opus46_inversion": {
|
| 225 |
+
"repo": "Jackrong/Claude-opus-4.6-TraceInversion-9000x",
|
| 226 |
+
"config": null,
|
| 227 |
+
"kind": "other",
|
| 228 |
+
"rights": "claude",
|
| 229 |
+
"weight": 0.7
|
| 230 |
+
},
|
| 231 |
+
"claude_opus47_inversion": {
|
| 232 |
+
"repo": "Jackrong/Claude-opus-4.7-TraceInversion-5000x",
|
| 233 |
+
"config": null,
|
| 234 |
+
"kind": "other",
|
| 235 |
+
"rights": "claude",
|
| 236 |
+
"weight": 0.7
|
| 237 |
+
},
|
| 238 |
+
"claude_opus48_max": {
|
| 239 |
+
"repo": "ronaldcmz/claude_opus_4.8_max_thinking_5k_v2",
|
| 240 |
+
"config": null,
|
| 241 |
+
"kind": "other",
|
| 242 |
+
"rights": "claude",
|
| 243 |
+
"weight": 0.45
|
| 244 |
+
},
|
| 245 |
+
"high_coder": {
|
| 246 |
+
"repo": "Crownelius/High-Coder-Reasoning-Multi-Turn",
|
| 247 |
+
"config": null,
|
| 248 |
+
"kind": "other",
|
| 249 |
+
"rights": "hunter",
|
| 250 |
+
"weight": 0.8
|
| 251 |
+
}
|
| 252 |
+
},
|
| 253 |
+
"ratios": {
|
| 254 |
+
"minicpm_core": 0.35,
|
| 255 |
+
"other_sft": 0.3,
|
| 256 |
+
"agent": 0.3,
|
| 257 |
+
"identity": 0.05
|
| 258 |
+
},
|
| 259 |
+
"quality_filter": "no-echo-no-loop-whole-target-v1"
|
| 260 |
+
},
|
| 261 |
+
"phase5_parent": "checkpoints/phase4-step-000900",
|
| 262 |
+
"phase5_source_positions": {
|
| 263 |
+
"minicpm_agent_code": 7606,
|
| 264 |
+
"minicpm_agent_general": 33957,
|
| 265 |
+
"minicpm_agent_search": 722,
|
| 266 |
+
"hermes_multi": 372,
|
| 267 |
+
"hermes_single": 318,
|
| 268 |
+
"openai_schema_multi": 331,
|
| 269 |
+
"smoltalk": 565,
|
| 270 |
+
"openr1_verified": 3023,
|
| 271 |
+
"glm_reasoning": 3397,
|
| 272 |
+
"gptoss_reasoning": 945,
|
| 273 |
+
"claude_opus46_inversion": 1587,
|
| 274 |
+
"claude_opus47_inversion": 1593,
|
| 275 |
+
"claude_opus48_max": 285600,
|
| 276 |
+
"high_coder": 478446
|
| 277 |
+
},
|
| 278 |
+
"phase5_source_stats": {
|
| 279 |
+
"minicpm_agent_code": {
|
| 280 |
+
"accepted": 379,
|
| 281 |
+
"rejected": {
|
| 282 |
+
"agent_schema": 5380,
|
| 283 |
+
"holdout": 380,
|
| 284 |
+
"agent_invalid": 1467
|
| 285 |
+
}
|
| 286 |
+
},
|
| 287 |
+
"minicpm_agent_general": {
|
| 288 |
+
"accepted": 999,
|
| 289 |
+
"rejected": {
|
| 290 |
+
"agent_schema": 30969,
|
| 291 |
+
"holdout": 1652,
|
| 292 |
+
"agent_invalid": 335,
|
| 293 |
+
"whole_sample_over_context": 3,
|
| 294 |
+
"no_observed_tool_result": 2
|
| 295 |
+
}
|
| 296 |
+
},
|
| 297 |
+
"minicpm_agent_search": {
|
| 298 |
+
"accepted": 300,
|
| 299 |
+
"rejected": {
|
| 300 |
+
"no_observed_tool_result": 330,
|
| 301 |
+
"holdout": 38,
|
| 302 |
+
"agent_invalid": 54
|
| 303 |
+
}
|
| 304 |
+
},
|
| 305 |
+
"hermes_multi": {
|
| 306 |
+
"accepted": 339,
|
| 307 |
+
"rejected": {
|
| 308 |
+
"holdout": 26,
|
| 309 |
+
"agent_invalid": 7
|
| 310 |
+
}
|
| 311 |
+
},
|
| 312 |
+
"hermes_single": {
|
| 313 |
+
"accepted": 296,
|
| 314 |
+
"rejected": {
|
| 315 |
+
"holdout": 15,
|
| 316 |
+
"agent_invalid": 7
|
| 317 |
+
}
|
| 318 |
+
},
|
| 319 |
+
"openai_schema_multi": {
|
| 320 |
+
"accepted": 309,
|
| 321 |
+
"rejected": {
|
| 322 |
+
"holdout": 22
|
| 323 |
+
}
|
| 324 |
+
},
|
| 325 |
+
"smoltalk": {
|
| 326 |
+
"accepted": 541,
|
| 327 |
+
"rejected": {
|
| 328 |
+
"holdout": 24
|
| 329 |
+
}
|
| 330 |
+
},
|
| 331 |
+
"openr1_verified": {
|
| 332 |
+
"accepted": 1771,
|
| 333 |
+
"rejected": {
|
| 334 |
+
"unusable_sft_row": 1107,
|
| 335 |
+
"whole_sample_over_context": 59,
|
| 336 |
+
"holdout": 145,
|
| 337 |
+
"prompt_echo": 3
|
| 338 |
+
}
|
| 339 |
+
},
|
| 340 |
+
"glm_reasoning": {
|
| 341 |
+
"accepted": 3222,
|
| 342 |
+
"rejected": {
|
| 343 |
+
"holdout": 168,
|
| 344 |
+
"whole_sample_over_context": 146,
|
| 345 |
+
"repetitive": 1,
|
| 346 |
+
"unusable_sft_row": 7,
|
| 347 |
+
"prompt_echo": 5
|
| 348 |
+
}
|
| 349 |
+
},
|
| 350 |
+
"gptoss_reasoning": {
|
| 351 |
+
"accepted": 886,
|
| 352 |
+
"rejected": {
|
| 353 |
+
"whole_sample_over_context": 357,
|
| 354 |
+
"holdout": 57,
|
| 355 |
+
"too_shallow_for_reasoning_mix": 86,
|
| 356 |
+
"prompt_echo": 27,
|
| 357 |
+
"generic_template": 1,
|
| 358 |
+
"unusable_sft_row": 2
|
| 359 |
+
}
|
| 360 |
+
},
|
| 361 |
+
"claude_opus46_inversion": {
|
| 362 |
+
"accepted": 1498,
|
| 363 |
+
"rejected": {
|
| 364 |
+
"holdout": 86,
|
| 365 |
+
"whole_sample_over_context": 5,
|
| 366 |
+
"repetitive": 3,
|
| 367 |
+
"unusable_sft_row": 3
|
| 368 |
+
}
|
| 369 |
+
},
|
| 370 |
+
"claude_opus47_inversion": {
|
| 371 |
+
"accepted": 1498,
|
| 372 |
+
"rejected": {
|
| 373 |
+
"holdout": 92,
|
| 374 |
+
"prompt_echo": 1,
|
| 375 |
+
"unusable_sft_row": 3,
|
| 376 |
+
"repetitive": 2
|
| 377 |
+
}
|
| 378 |
+
},
|
| 379 |
+
"claude_opus48_max": {
|
| 380 |
+
"accepted": 0,
|
| 381 |
+
"rejected": {
|
| 382 |
+
"generic_template": 266583,
|
| 383 |
+
"unusable_sft_row": 269611,
|
| 384 |
+
"holdout": 15989,
|
| 385 |
+
"prompt_echo": 3028
|
| 386 |
+
}
|
| 387 |
+
},
|
| 388 |
+
"high_coder": {
|
| 389 |
+
"accepted": 18,
|
| 390 |
+
"rejected": {
|
| 391 |
+
"no_visible_reasoning": 1367037,
|
| 392 |
+
"unusable_sft_row": 455667,
|
| 393 |
+
"too_shallow_for_reasoning_mix": 68262,
|
| 394 |
+
"holdout": 22761,
|
| 395 |
+
"whole_sample_over_context": 10
|
| 396 |
+
}
|
| 397 |
+
}
|
| 398 |
+
},
|
| 399 |
+
"identity": "Orion T2.1 2B",
|
| 400 |
+
"created_by": "Project Prism",
|
| 401 |
+
"source_checkpoint_repo": "Project-Prism/Orion-Flagship-2B-T2.2-Instruct-AgentML",
|
| 402 |
+
"source_checkpoint_folder": "checkpoints/phase4-step-000900",
|
| 403 |
+
"agent_format": "Orion AgentML: supervised tool calls, observed outputs, and final answers",
|
| 404 |
+
"warning": "Training steps do not establish reasoning quality; run independent evaluations."
|
| 405 |
+
}
|
checkpoints/phase5-step-004900/model-0000.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:59724a2e4f8d76ec9ab76ef3071fca86da391f24c3eeb3b6e639c639dab33928
|
| 3 |
+
size 366044336
|
checkpoints/phase5-step-004900/model-0001.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0bd5f7bad129c0cf67ac71d292f86ce519efad03f14107242c60ec3762020736
|
| 3 |
+
size 356846624
|
checkpoints/phase5-step-004900/model-0002.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:09e163d820cf808ef5026b1467ec46517e055b95fa020885f36f71717a10186d
|
| 3 |
+
size 356846656
|
checkpoints/phase5-step-004900/model-0003.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8007324f9e1eba97dbc63deb27042e21ba6e9f6a77948f68d5c322653056e5c0
|
| 3 |
+
size 355334276
|
checkpoints/phase5-step-004900/model-0004.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c3ddb6dc8f501c1e5b66cfbe62bf5c0ebbe0a460cc3e44e4fcc58a96bd557b38
|
| 3 |
+
size 350356824
|
checkpoints/phase5-step-004900/model-0005.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f8e3cae093f450d219e4be3813771458aeebc412535a55a7f874bb5aaa1a5acc
|
| 3 |
+
size 350356824
|
checkpoints/phase5-step-004900/model-0006.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6091c27be16f5b133183a88fc810fc8682fecf79e696a9ffb2b212fc68242c17
|
| 3 |
+
size 353719824
|
checkpoints/phase5-step-004900/model-0007.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2e9d89dc570c2da1745bca69779efe317df9c2f31f66fea668885351642eb2a9
|
| 3 |
+
size 350356824
|
checkpoints/phase5-step-004900/model-0008.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0b3de772b07f35c3d4907ca6dcc408b5ce5551036b63e687eab8f946d9303d6e
|
| 3 |
+
size 350356824
|
checkpoints/phase5-step-004900/model-0009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b55eb7f35e44e3f132442d6a4901374919b91f9e854c63dff08309ed80807d40
|
| 3 |
+
size 353719824
|
checkpoints/phase5-step-004900/model-0010.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:891c01dad0af84bea8c5613260ebfde8187ea6e0091644afdf552188db1a4f7d
|
| 3 |
+
size 350356824
|
checkpoints/phase5-step-004900/model-0011.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7bd234a60f2d1774411692c38dd06f002e40914e2d2604604f74659bfa6a5e14
|
| 3 |
+
size 350356824
|
checkpoints/phase5-step-004900/model-0012.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4c2c6aeb0ea645fcca6757dde51eb5551038ae01cfe95247d7f1e0d0833ceef1
|
| 3 |
+
size 353719824
|
checkpoints/phase5-step-004900/model-0013.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f55223a89d4831e7cd911015a7b9246c7d319691e06dd3a133e2484e2ce72935
|
| 3 |
+
size 350356824
|
checkpoints/phase5-step-004900/model-0014.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bce0270eb68e0d0126691f8a5aa1722fd08e8e6e3d3f75142e3e4fc905fcdf04
|
| 3 |
+
size 350356824
|
checkpoints/phase5-step-004900/model-0015.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7a23f83442e32811936c5c5dbc0746b473b5c079d92833de18596c010df6ef8c
|
| 3 |
+
size 361170384
|
checkpoints/phase5-step-004900/model-0016.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a4917a8d3a51262638c258776cbea520a9f6432dc8431b8c95d9d0ce73919b84
|
| 3 |
+
size 342906264
|
checkpoints/phase5-step-004900/model-0017.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0cadf383a27dad3d2606f21d56ae9c91a5b2b10e25e7caee329b4e6ec21c87e2
|
| 3 |
+
size 350356824
|
checkpoints/phase5-step-004900/model-0018.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2530b29bbc9f0d8d7556cd7edd8c6d065156bd7c75f4e0b9f6c0158991efee72
|
| 3 |
+
size 350356832
|
checkpoints/phase5-step-004900/model-0019.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ee6aa150971640a68447f24f1d09f38374a2016f4f2af21575028d0b96ef449b
|
| 3 |
+
size 353719816
|
checkpoints/phase5-step-004900/model-0020.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cee71897c8ecdd0d89cfc756ff0e4852f7725c3155276ea08639124636ed8020
|
| 3 |
+
size 350356824
|
checkpoints/phase5-step-004900/model-0021.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:99af3b42fe9c3745635313a05fce43ffdb5dbb8a69d2a668b53cbdd2e39976d2
|
| 3 |
+
size 350356832
|
checkpoints/phase5-step-004900/model-0022.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:bea1cfb0f43ad96e809b6f72fd41e579b3f153a4b190bbbeb7fef8d9bddc8553
|
| 3 |
+
size 355790520
|
checkpoints/phase5-step-004900/model-0023.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9b933416a727c5f17d04635536665ef3a1496e6dd16a1267e024120e3fa3d0f5
|
| 3 |
+
size 53746560
|
checkpoints/phase5-step-004900/orion_phase5_launcher.py
ADDED
|
@@ -0,0 +1,894 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import marimo
|
| 2 |
+
|
| 3 |
+
__generated_with = "0.25.1"
|
| 4 |
+
app = marimo.App(width="medium", auto_download=["html"])
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
@app.cell
|
| 8 |
+
def _():
|
| 9 |
+
def _():
|
| 10 |
+
import getpass
|
| 11 |
+
import os
|
| 12 |
+
from huggingface_hub import login
|
| 13 |
+
|
| 14 |
+
# Prompt for the token securely
|
| 15 |
+
hf_token = getpass.getpass("Enter your Hugging Face token: ").strip()
|
| 16 |
+
|
| 17 |
+
# Set it as an environment variable
|
| 18 |
+
os.environ["HF_TOKEN"] = hf_token
|
| 19 |
+
return
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
_()
|
| 23 |
+
return
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
@app.cell
|
| 27 |
+
def _():
|
| 28 |
+
#!/usr/bin/env python3
|
| 29 |
+
"""Orion T2.1 2B / T2.2 architecture β Phase 5 SFT, MoLab single-file launcher.
|
| 30 |
+
|
| 31 |
+
Requires original Orion Phase 4 trainer in your Hugging Face repository.
|
| 32 |
+
Uses its exact architecture, kernels, tokenizer, AgentML and safetensors format.
|
| 33 |
+
Run MODE='selftest' then 'inspect' then 'smoke' before 'train'.
|
| 34 |
+
Does not execute dataset tool calls. Does not train on generated hidden CoT.
|
| 35 |
+
"""
|
| 36 |
+
from __future__ import annotations
|
| 37 |
+
|
| 38 |
+
import ast
|
| 39 |
+
import collections
|
| 40 |
+
import getpass
|
| 41 |
+
import hashlib
|
| 42 |
+
import json
|
| 43 |
+
import math
|
| 44 |
+
import os
|
| 45 |
+
import random
|
| 46 |
+
import re
|
| 47 |
+
import shutil
|
| 48 |
+
import sys
|
| 49 |
+
import statistics
|
| 50 |
+
import time
|
| 51 |
+
from pathlib import Path
|
| 52 |
+
|
| 53 |
+
# ============================================================================
|
| 54 |
+
# OPERATOR SETTINGS β EDIT HERE. Start with selftest/inspect, NOT train.
|
| 55 |
+
# ============================================================================
|
| 56 |
+
class P5:
|
| 57 |
+
MODE = 'train' # selftest | inspect | token_audit | smoke | train | rescue
|
| 58 |
+
NAME = 'Orion T2.1 2B'
|
| 59 |
+
CREATOR = 'Project Prism'
|
| 60 |
+
ARCHITECTURE = 'orion_flagship_2b_t2_2'
|
| 61 |
+
SOURCE_REPO = 'Project-Prism/Orion-Flagship-2B-T2.2-Instruct-AgentML'
|
| 62 |
+
UPLOAD_REPO = SOURCE_REPO
|
| 63 |
+
SOURCE_TRAINER = 'checkpoints/phase4-step-000850/orion_unified_trainer.py'
|
| 64 |
+
TRAINER_REVISION = 'main' # Pin to a trusted commit SHA for reproducibility.
|
| 65 |
+
REQUIRE_PHASE4_COMPLETE = True # Refuse to train from step <900.
|
| 66 |
+
WORK_DIR = './orion_phase5_work'
|
| 67 |
+
STEPS = 5000
|
| 68 |
+
GRAD_ACCUM = 4
|
| 69 |
+
LEARNING_RATE = 8e-7
|
| 70 |
+
SAVE_EVERY = 100
|
| 71 |
+
LOG_EVERY = 10
|
| 72 |
+
EVAL_EVERY = 100
|
| 73 |
+
SEED = 520261011
|
| 74 |
+
KEEP_LOCAL = 2
|
| 75 |
+
MAX_CONTEXT = 32768
|
| 76 |
+
MIN_CONTEXT = 8192
|
| 77 |
+
MAX_SOURCE_SCAN = 300
|
| 78 |
+
REQUIRE_HUB_UPLOAD = True
|
| 79 |
+
SAVE_OPTIMIZER = False # True is large; false resets AdamW on resume.
|
| 80 |
+
# Real MiniCPM5 core SFT is gated. Request access on HF first.
|
| 81 |
+
REQUIRE_MINICPM_CORE = False
|
| 82 |
+
REQUIRE_MINICPM_AGENT = True
|
| 83 |
+
# Only turn on if YOU have checked upstream provider training rights.
|
| 84 |
+
ALLOW_OTHER_TEACHER_DATA = True
|
| 85 |
+
ALLOW_CLAUDE_TEACHER_DATA = True # TRAINING gate: require actual permission first.
|
| 86 |
+
PREVIEW_CLAUDE_SOURCES = True # Default ON for inspect/token_audit only; NEVER for training.
|
| 87 |
+
TOKEN_AUDIT_ROWS = 30 # Accepted samples per dataset; change to 100+ for better estimates.
|
| 88 |
+
ALLOW_HUNTER_TEACHER_DATA = True
|
| 89 |
+
# Core reasoning 35%, other reasoning/chat 30%, real agent 30%, identity 5%.
|
| 90 |
+
# When gated core unavailable, its share is reassigned to other chat/reasoning.
|
| 91 |
+
RATIOS = dict(minicpm_core=.35, other_sft=.30, agent=.30, identity=.05)
|
| 92 |
+
|
| 93 |
+
SOURCE_SPECS = {
|
| 94 |
+
'minicpm_core': dict(repo='openbmb/UltraData-SFT-2605', config=None,
|
| 95 |
+
kind='core', rights='public', weight=1.0),
|
| 96 |
+
'minicpm_agent_code': dict(repo='openbmb/UltraData-SFT-Agent-2609',
|
| 97 |
+
config='Code-Agent', kind='agent', rights='public', weight=1.0),
|
| 98 |
+
'minicpm_agent_general': dict(repo='openbmb/UltraData-SFT-Agent-2609',
|
| 99 |
+
config='General-Agent', kind='agent', rights='public', weight=1.2),
|
| 100 |
+
'minicpm_agent_search': dict(repo='openbmb/UltraData-SFT-Agent-2609',
|
| 101 |
+
config='Search-Agent', kind='agent', rights='public', weight=.8),
|
| 102 |
+
'hermes_multi': dict(repo='NousResearch/hermes-function-calling-v1',
|
| 103 |
+
config='func_calling', kind='agent', rights='public', weight=.6),
|
| 104 |
+
'hermes_single': dict(repo='NousResearch/hermes-function-calling-v1',
|
| 105 |
+
config='func_calling_singleturn', kind='agent', rights='public', weight=.35),
|
| 106 |
+
'openai_schema_multi': dict(repo='stindardlogic/openai-function-calling-5k',
|
| 107 |
+
config=None, kind='agent', rights='public', weight=.45),
|
| 108 |
+
'smoltalk': dict(repo='HuggingFaceTB/smoltalk', config='everyday-conversations',
|
| 109 |
+
kind='other', rights='public', weight=1.0),
|
| 110 |
+
'openr1_verified': dict(repo='open-r1/OpenR1-Math-220k', config='default',
|
| 111 |
+
kind='other', rights='public', weight=0.8),
|
| 112 |
+
'glm_reasoning': dict(repo='Jackrong/GLM-5.1-Reasoning-1M-Cleaned', config='main',
|
| 113 |
+
kind='other', rights='other', weight=1.5),
|
| 114 |
+
'gptoss_reasoning': dict(repo='Jackrong/ShareGPT-gpt-oss-120B-reasoning', config=None,
|
| 115 |
+
kind='other', rights='other', weight=1.0),
|
| 116 |
+
'claude_opus46_inversion': dict(repo='Jackrong/Claude-opus-4.6-TraceInversion-9000x', config=None,
|
| 117 |
+
kind='other', rights='claude', weight=.7),
|
| 118 |
+
'claude_opus47_inversion': dict(repo='Jackrong/Claude-opus-4.7-TraceInversion-5000x', config=None,
|
| 119 |
+
kind='other', rights='claude', weight=.7),
|
| 120 |
+
'claude_opus48_max': dict(repo='ronaldcmz/claude_opus_4.8_max_thinking_5k_v2', config=None,
|
| 121 |
+
kind='other', rights='claude', weight=.45),
|
| 122 |
+
'high_coder': dict(repo='Crownelius/High-Coder-Reasoning-Multi-Turn', config=None,
|
| 123 |
+
kind='other', rights='hunter', weight=.8),
|
| 124 |
+
}
|
| 125 |
+
|
| 126 |
+
# Hand-written, correct, non-repetitive identity + multi-turn examples.
|
| 127 |
+
IDENTITY = [
|
| 128 |
+
('Who are you?', 'I am Orion T2.1 2B, an experimental AI assistant built by Project Prism.'),
|
| 129 |
+
('What is your name?', 'My name is Orion T2.1 2B. You can call me Orion.'),
|
| 130 |
+
('Who made you?', 'Project Prism built me.'),
|
| 131 |
+
('Are you Claude or ChatGPT?', 'No. I am Orion T2.1 2B, a separate experimental model.'),
|
| 132 |
+
('How many parameters do you have?', 'About 2 billion parameters.'),
|
| 133 |
+
('What architecture are you?', 'My architecture is Orion Flagship 2B T2.2; my assistant name is Orion T2.1 2B.'),
|
| 134 |
+
('What is Super Mega Ultra Extreem?', 'It is my largest reasoning-effort setting, not a promise of correctness.'),
|
| 135 |
+
('Do you have to use all your thinking tokens?', 'No. I should stop reasoning once I have a sound answer.'),
|
| 136 |
+
('Hi', 'Hey! I am Orion. What can I help you with?'),
|
| 137 |
+
('Should you open test.py when someone says hello?', 'No. A greeting needs no tools.'),
|
| 138 |
+
('My favourite food is spaghetti. What is MY favourite food?', 'Your favourite food is spaghetti.'),
|
| 139 |
+
('I am worried you will make a mistake.', 'I can make mistakes. I will try to be clear and admit uncertainty.'),
|
| 140 |
+
]
|
| 141 |
+
MEMORY_CASES = [
|
| 142 |
+
(['My favourite colour is purple.', 'What is my favourite colour?'], 'Your favourite colour is purple.'),
|
| 143 |
+
(['My favourite food is spaghetti.', 'What is my favourite food?'], 'Your favourite food is spaghetti.'),
|
| 144 |
+
(['Your name is Orion. My name is Alex.', 'Who is Alex?'], 'Alex is the user, not me.'),
|
| 145 |
+
(['Remember: my pet is a cat named Pixel.', 'What is my cat called?'], 'Your cat is called Pixel.'),
|
| 146 |
+
]
|
| 147 |
+
# Small held-out diagnostic cases. These evaluate loss, NOT proof of reasoning.
|
| 148 |
+
EVAL_CASES = [
|
| 149 |
+
('A checkpoint is missing weight shards despite a success log. What should you do?',
|
| 150 |
+
'Do not load it. Verify the manifest and every shard, restore the last complete checkpoint, and use atomic saving with post-upload checks.'),
|
| 151 |
+
('All but nine sheep run away. How many remain?', 'Nine remain. "All but nine" means nine did not run away.'),
|
| 152 |
+
('What is 12 - 5 + 3?', '10. Subtract five from twelve to get seven, then add three.'),
|
| 153 |
+
('Should you use a file tool just to answer hello?', 'No. Say hello without opening any files.'),
|
| 154 |
+
]
|
| 155 |
+
|
| 156 |
+
|
| 157 |
+
def digest(obj):
|
| 158 |
+
return hashlib.sha256(json.dumps(obj, sort_keys=True, default=str, ensure_ascii=False).encode()).hexdigest()
|
| 159 |
+
|
| 160 |
+
|
| 161 |
+
def quality_check(question, answer, reason='', require_substance=True):
|
| 162 |
+
"""Cheap conservative filters, not an accuracy oracle."""
|
| 163 |
+
if not isinstance(question, str) or not isinstance(answer, str):
|
| 164 |
+
return 'non_string'
|
| 165 |
+
q, a = question.strip(), answer.strip()
|
| 166 |
+
if not q or not a or len(q) > 200000 or len(a) > 200000:
|
| 167 |
+
return 'empty_or_huge'
|
| 168 |
+
if a.casefold() == q.casefold() or (len(a) > 100 and q[:120].casefold() in a[:180].casefold()):
|
| 169 |
+
return 'prompt_echo'
|
| 170 |
+
if re.search(r'(?i)(?:i will deliver a structured|concrete next steps|analysis & recommendations)', a):
|
| 171 |
+
return 'generic_template'
|
| 172 |
+
# Detect repeated 6-word windows dominating an answer or a reasoning trace.
|
| 173 |
+
for text in (a, reason):
|
| 174 |
+
words = re.findall(r'\w+', text.casefold())
|
| 175 |
+
if len(words) >= 65:
|
| 176 |
+
sixes = collections.Counter(tuple(words[i:i+6]) for i in range(len(words)-5))
|
| 177 |
+
if sixes and sixes.most_common(1)[0][1] >= max(6, len(words)//15):
|
| 178 |
+
return 'repetitive'
|
| 179 |
+
if require_substance and len(a) < 55 and len(reason) < 100:
|
| 180 |
+
return 'too_shallow_for_reasoning_mix'
|
| 181 |
+
if re.search(r"(?i)\b(i am|i'm|my name is)\s+(claude|chatgpt|gpt-oss|glm)", a[:200]):
|
| 182 |
+
return 'wrong_identity'
|
| 183 |
+
return None
|
| 184 |
+
|
| 185 |
+
|
| 186 |
+
def load_orion_runtime():
|
| 187 |
+
"""Reuse the exact original 373-tensor trainer, without marimo UI startup."""
|
| 188 |
+
from huggingface_hub import hf_hub_download
|
| 189 |
+
path = hf_hub_download(repo_id=P5.SOURCE_REPO, filename=P5.SOURCE_TRAINER,
|
| 190 |
+
revision=P5.TRAINER_REVISION, token=os.environ.get('HF_TOKEN'))
|
| 191 |
+
source = Path(path).read_text(encoding='utf-8')
|
| 192 |
+
parsed = ast.parse(source)
|
| 193 |
+
cells = [n for n in ast.walk(parsed) if isinstance(n, ast.FunctionDef)
|
| 194 |
+
and n.name == '_' and any(isinstance(x, ast.ClassDef) and x.name == 'SFTConfig'
|
| 195 |
+
for x in n.body)]
|
| 196 |
+
if len(cells) != 1:
|
| 197 |
+
raise RuntimeError('Upstream trainer changed: cannot identify unique Orion SFT cell.')
|
| 198 |
+
cell = cells[0]
|
| 199 |
+
# Remove marimo-specific return and the old Phase4 entrypoint.
|
| 200 |
+
body = [n for n in cell.body if not isinstance(n, ast.Return) and not (
|
| 201 |
+
isinstance(n, ast.If) and isinstance(n.test, ast.Compare) and
|
| 202 |
+
isinstance(n.test.left, ast.Name) and n.test.left.id == '__name__')]
|
| 203 |
+
tree = ast.fix_missing_locations(ast.Module(body=body, type_ignores=[]))
|
| 204 |
+
ns = {'__name__': 'orion_phase5_runtime', '__file__': str(path)}
|
| 205 |
+
exec(compile(tree, str(path), 'exec'), ns)
|
| 206 |
+
if ns.get('P4_MODEL_ID') != P5.ARCHITECTURE or ns.get('P4_IDENTITY_NAME') != P5.NAME:
|
| 207 |
+
raise RuntimeError('Upstream architecture or Orion identity changed; review before training.')
|
| 208 |
+
for key in ('build_model', 'restore_model_shards', 'p3_hub_restore',
|
| 209 |
+
'save_checkpoint', 'p4_pack', 'parse_openai_dialogue', 'event_samples'):
|
| 210 |
+
if key not in ns:
|
| 211 |
+
raise RuntimeError('Upstream Orion trainer is missing '+key)
|
| 212 |
+
ns['CHECKPOINT_FORMATS'] += ('orion-phase5-1',)
|
| 213 |
+
ns['SFTConfig'].max_target_tokens = 32768
|
| 214 |
+
ns['SFTConfig'].max_prompt_tokens = 32768
|
| 215 |
+
ns['SFTConfig'].phase4_max_context = P5.MAX_CONTEXT
|
| 216 |
+
ns['P4_IDENTITY_NAME'] = P5.NAME
|
| 217 |
+
# Agent samples must also learn their identity; tools are contextual, not a command.
|
| 218 |
+
old_formatter = ns['format_context']
|
| 219 |
+
def with_identity(tools, events, closing_assistant=True):
|
| 220 |
+
s = old_formatter(tools, events, closing_assistant=closing_assistant)
|
| 221 |
+
return s.replace(ns['AGENTML_HEADER'], ns['AGENTML_HEADER'] + '\n' +
|
| 222 |
+
'You are Orion T2.1 2B, built by Project Prism. Use tools only when needed. '
|
| 223 |
+
'Never invent tool observations.\n', 1)
|
| 224 |
+
ns['format_context'] = with_identity
|
| 225 |
+
return ns
|
| 226 |
+
|
| 227 |
+
|
| 228 |
+
def allowed(spec):
|
| 229 |
+
rights = spec['rights']
|
| 230 |
+
return rights == 'public' or (rights == 'other' and P5.ALLOW_OTHER_TEACHER_DATA) or (
|
| 231 |
+
rights == 'claude' and P5.ALLOW_CLAUDE_TEACHER_DATA) or (
|
| 232 |
+
rights == 'hunter' and P5.ALLOW_HUNTER_TEACHER_DATA)
|
| 233 |
+
|
| 234 |
+
|
| 235 |
+
def mini_agent_row(row, ns):
|
| 236 |
+
"""Normalize MiniCPM messages for Orion; NEVER train on foreign system prompts.
|
| 237 |
+
|
| 238 |
+
Keep original user/assistant/tool chronology, real observations and tool-call
|
| 239 |
+
arguments. The existing Orion parsers decide whether an example is valid.
|
| 240 |
+
"""
|
| 241 |
+
original = ns['parse_json'](row.get('messages'))
|
| 242 |
+
if not isinstance(original, list):
|
| 243 |
+
return None
|
| 244 |
+
clean = []
|
| 245 |
+
for m in original:
|
| 246 |
+
if not isinstance(m, dict):
|
| 247 |
+
continue
|
| 248 |
+
role = str(m.get('role','')).lower()
|
| 249 |
+
if role in ('system', 'developer'):
|
| 250 |
+
continue # Includes "You are Codex" / OpenClaw behavioural instructions.
|
| 251 |
+
if role not in ('user', 'human', 'assistant', 'tool', 'function'):
|
| 252 |
+
continue
|
| 253 |
+
m = dict(m)
|
| 254 |
+
if role == 'human': m['role'] = 'user'
|
| 255 |
+
if role == 'function': m['role'] = 'tool'
|
| 256 |
+
clean.append(m)
|
| 257 |
+
if not clean or not any(m['role']=='user' for m in clean):
|
| 258 |
+
return None
|
| 259 |
+
tools = ns['parse_json'](row.get('tools'))
|
| 260 |
+
if not isinstance(tools, list): tools = []
|
| 261 |
+
return {'messages':clean, 'tools':tools}
|
| 262 |
+
|
| 263 |
+
|
| 264 |
+
def parse_minicpm_agent(row, ns):
|
| 265 |
+
"""Try native Orion parser then its schema-inference SWE-style fallback."""
|
| 266 |
+
clean = mini_agent_row(row, ns)
|
| 267 |
+
if clean is None: return None
|
| 268 |
+
# Different loaders accept different row shapes; do not invent tool results.
|
| 269 |
+
swe_row = {'messages': json.dumps(clean['messages'], ensure_ascii=False),
|
| 270 |
+
'model': 'orion_source_adapter', 'resolved': True}
|
| 271 |
+
parsers = (lambda:ns['parse_openai_dialogue'](clean),
|
| 272 |
+
lambda:ns['parse_source_row']('swesmith_claude', clean),
|
| 273 |
+
lambda:ns['parse_source_row']('swesmith_claude', swe_row))
|
| 274 |
+
for parser in parsers:
|
| 275 |
+
try:
|
| 276 |
+
parsed = parser()
|
| 277 |
+
if parsed is not None and len(parsed)==3:
|
| 278 |
+
return parsed
|
| 279 |
+
except (TypeError, ValueError, KeyError, AttributeError):
|
| 280 |
+
continue
|
| 281 |
+
return None
|
| 282 |
+
|
| 283 |
+
|
| 284 |
+
def minicpm_code_features(ds):
|
| 285 |
+
"""Add the optional positional tool field found in Code-Agent Parquet shards.
|
| 286 |
+
|
| 287 |
+
Some shards contain this field while the repository's inferred Features omit it,
|
| 288 |
+
producing the observed Arrow cast error on iteration.
|
| 289 |
+
"""
|
| 290 |
+
import copy
|
| 291 |
+
from datasets import Value, List
|
| 292 |
+
features = copy.deepcopy(ds.features)
|
| 293 |
+
if features is None or 'tools' not in features:
|
| 294 |
+
return None
|
| 295 |
+
tool_feature = features['tools']
|
| 296 |
+
entry = getattr(tool_feature,'feature',None)
|
| 297 |
+
if isinstance(entry,dict) and 'positional' not in entry:
|
| 298 |
+
entry['positional'] = List(Value('string'))
|
| 299 |
+
return features
|
| 300 |
+
return None
|
| 301 |
+
|
| 302 |
+
|
| 303 |
+
def minicpm_general_chat(row, ns):
|
| 304 |
+
"""Strict text-only General-Agent fallback; no pretending to have tools."""
|
| 305 |
+
clean = mini_agent_row(row, ns)
|
| 306 |
+
if clean is None:
|
| 307 |
+
return None
|
| 308 |
+
messages = clean['messages']
|
| 309 |
+
# A transcript that ever uses tools MUST go through the AgentML parser.
|
| 310 |
+
if any(m.get('role') == 'tool' or m.get('tool_calls') or m.get('recipient')
|
| 311 |
+
or m.get('function_call') for m in messages):
|
| 312 |
+
return None
|
| 313 |
+
# Only accept an actual assistant final following a real user request.
|
| 314 |
+
history, answer, reason = [], None, ''
|
| 315 |
+
for m in messages:
|
| 316 |
+
role = m.get('role')
|
| 317 |
+
body = m.get('content', '')
|
| 318 |
+
if isinstance(body, list):
|
| 319 |
+
body = '\n'.join(str(x.get('text', '')) for x in body if isinstance(x, dict))
|
| 320 |
+
if not isinstance(body, str):
|
| 321 |
+
return None
|
| 322 |
+
if role == 'user' and body.strip():
|
| 323 |
+
history.append({'role': 'user', 'content': body.strip()})
|
| 324 |
+
answer = None
|
| 325 |
+
elif role == 'assistant' and history and body.strip():
|
| 326 |
+
answer = body.strip()
|
| 327 |
+
reason = str(m.get('reasoning_content') or m.get('thinking') or '')
|
| 328 |
+
history.append({'role': 'assistant', 'content': answer})
|
| 329 |
+
if not answer or not history or history[-1]['role'] != 'assistant':
|
| 330 |
+
return None
|
| 331 |
+
q = next((m['content'] for m in reversed(history[:-1]) if m['role']=='user'), '')
|
| 332 |
+
if quality_check(q, answer, reason, require_substance=False):
|
| 333 |
+
return None
|
| 334 |
+
return {'history': history[:-1], 'answer': answer, 'reason': reason,
|
| 335 |
+
'source': 'minicpm_agent_general_chat'}
|
| 336 |
+
|
| 337 |
+
|
| 338 |
+
class SourceStream:
|
| 339 |
+
def __init__(self, key, ns, offset=0):
|
| 340 |
+
self.key, self.ns = key, ns
|
| 341 |
+
self.spec = SOURCE_SPECS[key]
|
| 342 |
+
self.seed = P5.SEED + sum(map(ord, key))
|
| 343 |
+
self.raw_rows = 0
|
| 344 |
+
self.accepted = 0
|
| 345 |
+
self.rejected = collections.Counter()
|
| 346 |
+
self.pending = []
|
| 347 |
+
self.open()
|
| 348 |
+
if offset:
|
| 349 |
+
print(f'β©οΈ Replaying {offset} rows from {key} for approximate resume', flush=True)
|
| 350 |
+
for _ in range(offset): self.raw()
|
| 351 |
+
|
| 352 |
+
def open(self):
|
| 353 |
+
args = dict(path=self.spec['repo'], name=self.spec['config'],
|
| 354 |
+
split='train', streaming=True, token=os.environ.get('HF_TOKEN'))
|
| 355 |
+
ds = self.ns['load_dataset'](**args)
|
| 356 |
+
if self.key == 'minicpm_agent_code':
|
| 357 |
+
revised = minicpm_code_features(ds)
|
| 358 |
+
if revised is not None:
|
| 359 |
+
# Re-open with the complete nested schema BEFORE streaming rows.
|
| 360 |
+
# A cast after loading can fail before it has a chance to map.
|
| 361 |
+
try:
|
| 362 |
+
ds = self.ns['load_dataset'](**dict(args,features=revised))
|
| 363 |
+
print('π οΈ Code-Agent: positional tool schema included',flush=True)
|
| 364 |
+
except (TypeError, ValueError) as exc:
|
| 365 |
+
print('β οΈ Code-Agent schema override unavailable:',exc,flush=True)
|
| 366 |
+
self.iterator = iter(ds.shuffle(seed=self.seed, buffer_size=128))
|
| 367 |
+
print(f'π {self.key}: {self.spec["repo"]} / {self.spec["config"]}', flush=True)
|
| 368 |
+
|
| 369 |
+
def raw(self):
|
| 370 |
+
try: row = next(self.iterator)
|
| 371 |
+
except StopIteration:
|
| 372 |
+
self.seed += 1
|
| 373 |
+
self.open()
|
| 374 |
+
row = next(self.iterator)
|
| 375 |
+
self.raw_rows += 1
|
| 376 |
+
return row
|
| 377 |
+
|
| 378 |
+
def next_case(self, tokenizer=None):
|
| 379 |
+
"""Return one actual, filtered case; never synthesize missing answers."""
|
| 380 |
+
ns = self.ns
|
| 381 |
+
if self.pending: return self.pending.pop(0)
|
| 382 |
+
for _ in range(P5.MAX_SOURCE_SCAN):
|
| 383 |
+
row = self.raw()
|
| 384 |
+
# Stable 5% holdout partition. No holdout samples are used for training.
|
| 385 |
+
if int(digest(row)[:8], 16) % 20 == 0:
|
| 386 |
+
self.rejected['holdout'] += 1
|
| 387 |
+
continue
|
| 388 |
+
if self.spec['kind'] == 'agent':
|
| 389 |
+
# Actual MiniCPM5 tool schema: messages + tools, often JSON cells.
|
| 390 |
+
row = dict(row)
|
| 391 |
+
row['messages'] = ns['parse_json'](row.get('messages'))
|
| 392 |
+
row['tools'] = ns['parse_json'](row.get('tools'))
|
| 393 |
+
if self.key == 'hermes_multi':
|
| 394 |
+
parsed = ns['parse_source_row']('hermes_multi', row)
|
| 395 |
+
elif self.key == 'hermes_single':
|
| 396 |
+
parsed = ns['parse_source_row']('hermes', row)
|
| 397 |
+
elif self.key == 'openai_schema_multi':
|
| 398 |
+
parsed = ns['parse_source_row']('openai_multiturn', row)
|
| 399 |
+
else:
|
| 400 |
+
parsed = parse_minicpm_agent(row, ns)
|
| 401 |
+
if parsed is None:
|
| 402 |
+
# General-Agent includes both pure chat and tool tasks. Keep
|
| 403 |
+
# genuine text-only responses as CHAT, not fabricated agent
|
| 404 |
+
# examples. Reject partially observed tool conversations.
|
| 405 |
+
if self.key == 'minicpm_agent_general':
|
| 406 |
+
fallback = minicpm_general_chat(row, ns)
|
| 407 |
+
if fallback is not None:
|
| 408 |
+
self.accepted += 1
|
| 409 |
+
return 'chat', fallback
|
| 410 |
+
self.rejected['agent_schema'] += 1
|
| 411 |
+
continue
|
| 412 |
+
tools, events, _ = parsed
|
| 413 |
+
if ns['validate_events'](tools, events):
|
| 414 |
+
self.rejected['agent_invalid'] += 1
|
| 415 |
+
continue
|
| 416 |
+
if not any(e['role'] == 'observation' for e in events):
|
| 417 |
+
# Single-turn Hermes examples teach *requesting* tools. They
|
| 418 |
+
# have no results by design; never fabricate observations.
|
| 419 |
+
if self.key != 'hermes_single':
|
| 420 |
+
self.rejected['no_observed_tool_result'] += 1
|
| 421 |
+
continue
|
| 422 |
+
samples = ns['event_samples'](tools, events, tokenizer=tokenizer)
|
| 423 |
+
if not samples:
|
| 424 |
+
self.rejected['no_supervised_actions'] += 1
|
| 425 |
+
continue
|
| 426 |
+
# Prefer an action or final answer AFTER a real tool observation.
|
| 427 |
+
samples.sort(key=lambda s: ('<|observation|>' in s[0], s[2]), reverse=True)
|
| 428 |
+
self.accepted += 1
|
| 429 |
+
self.pending = [('agent', x) for x in samples[1:4]]
|
| 430 |
+
return 'agent', samples[0]
|
| 431 |
+
# Reuse Orion's existing multi-turn and <think> adapters.
|
| 432 |
+
adapter = self.key if self.key in ns['P4_DATA'] else 'smoltalk_everyday'
|
| 433 |
+
if self.key == 'claude_opus46_inversion': adapter = 'opus46_inversion'
|
| 434 |
+
if self.key == 'claude_opus47_inversion': adapter = 'opus47_inversion'
|
| 435 |
+
if self.key == 'claude_opus48_max': adapter = 'opus48_max'
|
| 436 |
+
if self.key == 'glm_reasoning': adapter = 'glm51_cleaned'
|
| 437 |
+
if self.key == 'gptoss_reasoning': adapter = 'gptoss_sharegpt'
|
| 438 |
+
if self.key == 'smoltalk': adapter = 'smoltalk_everyday'
|
| 439 |
+
if self.key == 'openr1_verified':
|
| 440 |
+
parsed = ns['p3_parse_openr1'](row)
|
| 441 |
+
cases = ([dict(history=[dict(role='user',content=parsed['question'])],
|
| 442 |
+
answer=parsed['answer'],reason=parsed['rationale'],source=self.key)] if parsed else [])
|
| 443 |
+
else:
|
| 444 |
+
cases = ns['p4_training_cases'](adapter, row)
|
| 445 |
+
good = []
|
| 446 |
+
for c in cases:
|
| 447 |
+
q = next((t['content'] for t in reversed(c['history']) if t['role'] == 'user'), '')
|
| 448 |
+
reason = c.get('reason', '')
|
| 449 |
+
issue = quality_check(q, c['answer'], reason, require_substance=(self.key != 'smoltalk'))
|
| 450 |
+
if issue:
|
| 451 |
+
self.rejected[issue] += 1
|
| 452 |
+
continue
|
| 453 |
+
if not reason and self.key not in ('smoltalk',):
|
| 454 |
+
# A long final answer is valid, but don't label it as a hidden trace.
|
| 455 |
+
self.rejected['no_visible_reasoning'] += 1
|
| 456 |
+
continue
|
| 457 |
+
good.append(c)
|
| 458 |
+
if good:
|
| 459 |
+
self.accepted += 1
|
| 460 |
+
self.pending = [('chat', c) for c in good[1:5]]
|
| 461 |
+
return 'chat', good[0]
|
| 462 |
+
self.rejected['unusable_sft_row'] += 1
|
| 463 |
+
raise RuntimeError(f'{self.key}: no usable samples in {P5.MAX_SOURCE_SCAN} rows; '
|
| 464 |
+
f'counts={dict(self.rejected)}. Inspect schema/permissions.')
|
| 465 |
+
|
| 466 |
+
|
| 467 |
+
def identity_case(ns, rng):
|
| 468 |
+
if rng.random() < .65:
|
| 469 |
+
q, a = rng.choice(IDENTITY)
|
| 470 |
+
history = [{'role':'user', 'content':q}]
|
| 471 |
+
else:
|
| 472 |
+
turns, a = rng.choice(MEMORY_CASES)
|
| 473 |
+
history = [{'role':'user','content':turns[0]},
|
| 474 |
+
{'role':'assistant','answer':'Got it.','reason':''},
|
| 475 |
+
{'role':'user','content':turns[1]}]
|
| 476 |
+
return dict(history=history, answer=a, reason='', effort='Instant', source='identity')
|
| 477 |
+
|
| 478 |
+
|
| 479 |
+
def pack_case(ns, category, value, tokenizer, cap, rng):
|
| 480 |
+
if category == 'agent':
|
| 481 |
+
prompt, target, final = value
|
| 482 |
+
# Context may be too large; skip whole examples rather than truncating.
|
| 483 |
+
ns['ACTIVE_CONTEXT_LIMIT'] = cap
|
| 484 |
+
sample = ns['agent_tensors']((prompt, target, final), tokenizer)
|
| 485 |
+
return sample
|
| 486 |
+
if category == 'identity':
|
| 487 |
+
value = identity_case(ns, rng)
|
| 488 |
+
reason = value.get('reason', '')
|
| 489 |
+
effort = value.get('effort') or (rng.choice(['High','Xhigh','Max']) if len(reason) > 400 else
|
| 490 |
+
rng.choice(['Low','Medium']) if reason else 'Instant')
|
| 491 |
+
sample, _ = ns['p4_pack'](value, tokenizer, effort, cap)
|
| 492 |
+
return sample
|
| 493 |
+
|
| 494 |
+
|
| 495 |
+
def cap_for_step(step):
|
| 496 |
+
f = step / P5.STEPS
|
| 497 |
+
return 8192 if f < .20 else 16384 if f < .55 else 32768
|
| 498 |
+
|
| 499 |
+
|
| 500 |
+
def fp(source_folder):
|
| 501 |
+
return dict(version='orion-phase5-minicpm-agent-v1', architecture=P5.ARCHITECTURE,
|
| 502 |
+
identity=P5.NAME, phase4_parent=source_folder, steps=P5.STEPS,
|
| 503 |
+
lr=P5.LEARNING_RATE, grad_accum=P5.GRAD_ACCUM,
|
| 504 |
+
max_context=P5.MAX_CONTEXT, sources={k:SOURCE_SPECS[k] for k in SOURCE_SPECS if allowed(SOURCE_SPECS[k])},
|
| 505 |
+
ratios=P5.RATIOS, quality_filter='no-echo-no-loop-whole-target-v1')
|
| 506 |
+
|
| 507 |
+
|
| 508 |
+
def latest_remote(ns, prefix):
|
| 509 |
+
files = ns['HfApi'](token=os.environ.get('HF_TOKEN')).list_repo_files(
|
| 510 |
+
repo_id=P5.SOURCE_REPO, repo_type='model')
|
| 511 |
+
hits = []
|
| 512 |
+
for f in files:
|
| 513 |
+
m = re.fullmatch(r'checkpoints/' + re.escape(prefix) + r'-step-(\d{6})/manifest\.json', f)
|
| 514 |
+
if m: hits.append((int(m.group(1)), f.rsplit('/', 1)[0]))
|
| 515 |
+
return sorted(hits, reverse=True)
|
| 516 |
+
|
| 517 |
+
|
| 518 |
+
def pick_checkpoint(ns, work):
|
| 519 |
+
local = []
|
| 520 |
+
for d in work.glob('phase5-step-*'):
|
| 521 |
+
try:
|
| 522 |
+
mf = ns['check_checkpoint_folder'](d, ('orion-phase5-1',))
|
| 523 |
+
if mf.get('phase5_fingerprint') == fp(mf['phase5_parent']):
|
| 524 |
+
local.append((int(mf['step']), d, mf))
|
| 525 |
+
except (ValueError, KeyError, OSError): pass
|
| 526 |
+
best = max(local, key=lambda x:x[0]) if local else None
|
| 527 |
+
for step, rel in latest_remote(ns, 'phase5'):
|
| 528 |
+
if best and step <= best[0]: break
|
| 529 |
+
try:
|
| 530 |
+
path = ns['hf_hub_download'](repo_id=P5.SOURCE_REPO,
|
| 531 |
+
filename=rel+'/manifest.json', token=os.environ.get('HF_TOKEN'))
|
| 532 |
+
meta = json.loads(Path(path).read_text())
|
| 533 |
+
if meta.get('phase5_fingerprint') != fp(meta.get('phase5_parent')):
|
| 534 |
+
print('β οΈ Skipping incompatible Phase5 checkpoint', rel, flush=True)
|
| 535 |
+
continue
|
| 536 |
+
folder, mf = ns['p3_hub_restore'](P5.SOURCE_REPO, rel, work,
|
| 537 |
+
('orion-phase5-1',), 'phase5', step)
|
| 538 |
+
best = (step, folder, mf)
|
| 539 |
+
break
|
| 540 |
+
except (OSError, ValueError, RuntimeError, FileNotFoundError) as exc:
|
| 541 |
+
print('β οΈ Skipping incomplete Phase5 checkpoint:', rel, exc, flush=True)
|
| 542 |
+
if best: return best
|
| 543 |
+
for step, rel in latest_remote(ns, 'phase4'):
|
| 544 |
+
if P5.REQUIRE_PHASE4_COMPLETE and step < 900:
|
| 545 |
+
raise RuntimeError(f'Latest Phase4 checkpoint is {step}, not 900. Finish Phase4 first.')
|
| 546 |
+
try:
|
| 547 |
+
folder, mf = ns['p3_hub_restore'](P5.SOURCE_REPO, rel, work,
|
| 548 |
+
('orion-phase4-1',), 'phase4', step)
|
| 549 |
+
return 0, folder, mf
|
| 550 |
+
except (OSError, ValueError, RuntimeError, FileNotFoundError) as exc:
|
| 551 |
+
print('β οΈ Skipping incomplete Phase4 checkpoint:', rel, exc, flush=True)
|
| 552 |
+
raise RuntimeError('No complete Phase4 source checkpoint was found.')
|
| 553 |
+
|
| 554 |
+
|
| 555 |
+
def select_streams(ns, offsets=None, for_inspect=False):
|
| 556 |
+
offsets = offsets or {}
|
| 557 |
+
active = {}
|
| 558 |
+
for key, spec in SOURCE_SPECS.items():
|
| 559 |
+
if not allowed(spec):
|
| 560 |
+
if not (for_inspect and spec['rights'] == 'claude' and P5.PREVIEW_CLAUDE_SOURCES):
|
| 561 |
+
print('π Training rights gate:', key, '(disabled)', flush=True)
|
| 562 |
+
continue
|
| 563 |
+
print('π CLAUDE PREVIEW ONLY:', key, '(NOT enabled for training)', flush=True)
|
| 564 |
+
try:
|
| 565 |
+
active[key] = SourceStream(key, ns, offset=0 if for_inspect else int(offsets.get(key,0)))
|
| 566 |
+
except Exception as exc:
|
| 567 |
+
print(f'β οΈ {key} unavailable: {type(exc).__name__}: {exc}', flush=True)
|
| 568 |
+
if P5.REQUIRE_MINICPM_AGENT and not any(k.startswith('minicpm_agent_') for k in active):
|
| 569 |
+
raise RuntimeError('MiniCPM5 agent mix required but no agent config could be opened.')
|
| 570 |
+
if P5.REQUIRE_MINICPM_CORE and 'minicpm_core' not in active:
|
| 571 |
+
raise RuntimeError('MiniCPM5 core SFT required but gated/unavailable. Request access on Hugging Face.')
|
| 572 |
+
if 'minicpm_core' not in active:
|
| 573 |
+
print('β οΈ MiniCPM5 CORE SFT unavailable (possibly gated). Other SFT sources replace its share.', flush=True)
|
| 574 |
+
if not any(s.spec['kind'] == 'other' for s in active.values()) and 'minicpm_core' not in active:
|
| 575 |
+
raise RuntimeError('No general SFT source available; refusing agent-only training.')
|
| 576 |
+
return active
|
| 577 |
+
|
| 578 |
+
|
| 579 |
+
def choose_sample(ns, streams, tokenizer, cap, rng):
|
| 580 |
+
bykind = collections.defaultdict(list)
|
| 581 |
+
for key, s in streams.items(): bykind[s.spec['kind']].append(s)
|
| 582 |
+
ratios = dict(P5.RATIOS)
|
| 583 |
+
if not bykind['core']:
|
| 584 |
+
ratios['other_sft'] += ratios.pop('minicpm_core')
|
| 585 |
+
categories = [('core',ratios.get('minicpm_core',0)), ('other',ratios['other_sft']),
|
| 586 |
+
('agent',ratios['agent']), ('identity',ratios['identity'])]
|
| 587 |
+
categories = [(k,v) for k,v in categories if v > 0 and (k == 'identity' or bykind[k])]
|
| 588 |
+
for _ in range(180):
|
| 589 |
+
group = rng.choices([k for k,v in categories],weights=[v for k,v in categories])[0]
|
| 590 |
+
if group == 'identity':
|
| 591 |
+
sample = pack_case(ns,'identity',None,tokenizer,cap,rng)
|
| 592 |
+
if sample is not None: return 'identity',sample
|
| 593 |
+
continue
|
| 594 |
+
sources = bykind[group]
|
| 595 |
+
stream = rng.choices(sources,weights=[s.spec['weight'] for s in sources])[0]
|
| 596 |
+
try:
|
| 597 |
+
kind, value = stream.next_case(tokenizer)
|
| 598 |
+
except Exception as exc:
|
| 599 |
+
print('β οΈ Source quarantined:', exc, flush=True)
|
| 600 |
+
bykind[group] = [s for s in bykind[group] if s is not stream]
|
| 601 |
+
if group == 'agent' and P5.REQUIRE_MINICPM_AGENT and not bykind[group]:
|
| 602 |
+
raise RuntimeError('All required MiniCPM5 agent sources failed; refusing silent fallback.') from exc
|
| 603 |
+
if group == 'core' and P5.REQUIRE_MINICPM_CORE and not bykind[group]:
|
| 604 |
+
raise RuntimeError('Required MiniCPM5 core source failed; check gated access.') from exc
|
| 605 |
+
categories = [(k,v) for k,v in categories if k != group or bykind[group]]
|
| 606 |
+
if not categories: raise RuntimeError('All sources quarantined')
|
| 607 |
+
continue
|
| 608 |
+
sample = pack_case(ns,kind,value,tokenizer,cap,rng)
|
| 609 |
+
if sample is not None: return stream.key, sample
|
| 610 |
+
stream.rejected['whole_sample_over_context'] += 1
|
| 611 |
+
raise RuntimeError('No whole sample could be packed after 180 attempts; inspect source quality.')
|
| 612 |
+
|
| 613 |
+
|
| 614 |
+
def eval_loss(ns, model, tok):
|
| 615 |
+
torch = ns['torch']
|
| 616 |
+
scores = []
|
| 617 |
+
model.eval()
|
| 618 |
+
try:
|
| 619 |
+
for q,a in EVAL_CASES:
|
| 620 |
+
c = dict(history=[dict(role='user',content=q)], answer=a, reason='')
|
| 621 |
+
sample, _ = ns['p4_pack'](c,tok,'Medium',8192)
|
| 622 |
+
if sample is None: continue
|
| 623 |
+
x,y = sample
|
| 624 |
+
with torch.no_grad(), torch.autocast(device_type='cuda',dtype=torch.bfloat16):
|
| 625 |
+
ce, _ = model(x.unsqueeze(0).to(ns['DEVICE']), y.unsqueeze(0).to(ns['DEVICE']))
|
| 626 |
+
scores.append(float(ce.detach()))
|
| 627 |
+
finally:
|
| 628 |
+
model.train()
|
| 629 |
+
return sum(scores)/len(scores) if scores else float('nan')
|
| 630 |
+
|
| 631 |
+
|
| 632 |
+
def save(ns, model, tok, opt, arch, step, streams, work, parent):
|
| 633 |
+
dest = work/f'phase5-step-{step:06d}'
|
| 634 |
+
if dest.exists(): raise FileExistsError(dest)
|
| 635 |
+
extra = dict(format_version='orion-phase5-1', training_stage='phase5',
|
| 636 |
+
phase5_fingerprint=fp(parent), phase5_parent=parent,
|
| 637 |
+
phase5_source_positions={k:s.raw_rows for k,s in streams.items()},
|
| 638 |
+
phase5_source_stats={k:dict(accepted=s.accepted,rejected=dict(s.rejected)) for k,s in streams.items()},
|
| 639 |
+
identity=P5.NAME, created_by=P5.CREATOR,
|
| 640 |
+
source_checkpoint_repo=P5.SOURCE_REPO, source_checkpoint_folder=parent,
|
| 641 |
+
agent_format='Orion AgentML: supervised tool calls, observed outputs, and final answers',
|
| 642 |
+
warning='Training steps do not establish reasoning quality; run independent evaluations.')
|
| 643 |
+
# Saving requires free disk for 24+ shards, never prune before verified upload.
|
| 644 |
+
estimate = sum(p.numel()*p.element_size() for p in model.parameters())
|
| 645 |
+
if shutil.disk_usage(work).free < estimate + 2*1024**3:
|
| 646 |
+
raise RuntimeError('Not enough disk space for an atomic Phase5 checkpoint.')
|
| 647 |
+
item = ns['save_checkpoint'](model,tok,opt,dest,arch,step,
|
| 648 |
+
{k:s.raw_rows for k,s in streams.items()},save_optimizer=P5.SAVE_OPTIMIZER,
|
| 649 |
+
manifest_overrides=extra)
|
| 650 |
+
ns['check_checkpoint_folder'](item,('orion-phase5-1',))
|
| 651 |
+
shutil.copy2(Path(__file__), item/'orion_phase5_launcher.py')
|
| 652 |
+
if P5.REQUIRE_HUB_UPLOAD:
|
| 653 |
+
ns['upload_checkpoint'](item,P5.UPLOAD_REPO)
|
| 654 |
+
ns['combo_uploaded_files'](item,P5.UPLOAD_REPO)
|
| 655 |
+
folders = sorted(work.glob('phase5-step-*'))
|
| 656 |
+
for old in folders[:-max(2,P5.KEEP_LOCAL)]:
|
| 657 |
+
if old != item and old.is_dir():
|
| 658 |
+
ns['check_checkpoint_folder'](old,('orion-phase5-1',))
|
| 659 |
+
shutil.rmtree(old)
|
| 660 |
+
return item
|
| 661 |
+
|
| 662 |
+
|
| 663 |
+
def selftest():
|
| 664 |
+
assert sum(P5.RATIOS.values()) == 1
|
| 665 |
+
assert quality_check('What is your name?', 'What is your name?') == 'prompt_echo'
|
| 666 |
+
assert quality_check('Hi', 'Hello!', require_substance=False) is None
|
| 667 |
+
assert quality_check('A?', 'is not a sheep that ran away '*40, require_substance=False) == 'repetitive'
|
| 668 |
+
assert quality_check('Question', 'Here is a detailed, nonrepeating explanation that answers the question clearly.',
|
| 669 |
+
require_substance=False) is None
|
| 670 |
+
assert cap_for_step(1) == 8192 and cap_for_step(P5.STEPS) == 32768
|
| 671 |
+
assert P5.ARCHITECTURE == 'orion_flagship_2b_t2_2'
|
| 672 |
+
assert all(k in SOURCE_SPECS for k in ('claude_opus46_inversion', 'claude_opus47_inversion', 'claude_opus48_max'))
|
| 673 |
+
if not P5.ALLOW_CLAUDE_TEACHER_DATA:
|
| 674 |
+
assert not allowed(SOURCE_SPECS['claude_opus47_inversion'])
|
| 675 |
+
assert P5.PREVIEW_CLAUDE_SOURCES is True
|
| 676 |
+
fake_ns = {'parse_json':lambda x:x}
|
| 677 |
+
mini = mini_agent_row({'messages':[
|
| 678 |
+
{'role':'system','content':'You are Codex, not Orion.'},
|
| 679 |
+
{'role':'user','content':'Read file'},
|
| 680 |
+
{'role':'assistant','content':'','tool_calls':[{'type':'function','function':{'name':'read','arguments':{'path':'a'}}}]},
|
| 681 |
+
{'role':'tool','content':'File contents'}], 'tools':[]},fake_ns)
|
| 682 |
+
assert mini and len(mini['messages']) == 3
|
| 683 |
+
assert not any('You are Codex' in str(m) for m in mini['messages'])
|
| 684 |
+
assert mini['messages'][1]['tool_calls'][0]['function']['name'] == 'read'
|
| 685 |
+
assert mini['messages'][2]['role'] == 'tool'
|
| 686 |
+
plain = {'messages': [
|
| 687 |
+
{'role':'system','content':'You are Claude'},
|
| 688 |
+
{'role':'user','content':'Explain a useful concept in plain English.'},
|
| 689 |
+
{'role':'assistant','content':'A good explanation uses a simple definition, an example, and one clear limitation.'}]}
|
| 690 |
+
case = minicpm_general_chat(plain, fake_ns)
|
| 691 |
+
assert case and case['history'][0]['role']=='user'
|
| 692 |
+
assert not minicpm_general_chat({'messages': plain['messages'] +
|
| 693 |
+
[{'role':'tool','content':'unverified'}]}, fake_ns)
|
| 694 |
+
print('β
Phase5 local CPU selftest passed. No downloads, no GPU, no training.')
|
| 695 |
+
|
| 696 |
+
|
| 697 |
+
def inspect(ns):
|
| 698 |
+
streams = select_streams(ns, for_inspect=True)
|
| 699 |
+
print('\nπ SOURCE AUDIT (real rows, never used for training in this mode)')
|
| 700 |
+
for name, s in streams.items():
|
| 701 |
+
for _ in range(2):
|
| 702 |
+
try:
|
| 703 |
+
kind, case = s.next_case()
|
| 704 |
+
if kind == 'agent':
|
| 705 |
+
prompt,target,final=case
|
| 706 |
+
print(f' {name}: agent final={final} context={len(prompt)} chars target={len(target)} chars')
|
| 707 |
+
else:
|
| 708 |
+
print(f' {name}: chat answer={len(case["answer"])} chars '
|
| 709 |
+
f'reason={len(case.get("reason", ""))} chars history={len(case["history"])} turns')
|
| 710 |
+
except Exception as exc:
|
| 711 |
+
print(f' {name}: β οΈ {type(exc).__name__}: {exc}')
|
| 712 |
+
break
|
| 713 |
+
print(f' accepted={s.accepted}, rejected={dict(s.rejected)}')
|
| 714 |
+
print('Inspect sample acceptance and data rights before switching MODE to smoke/train.')
|
| 715 |
+
|
| 716 |
+
|
| 717 |
+
def token_audit(ns):
|
| 718 |
+
"""Measure REAL Mistral-tokenizer lengths of accepted, complete examples.
|
| 719 |
+
|
| 720 |
+
Sampling is approximate and reports only accepted rows; no GPU/weights/training.
|
| 721 |
+
Claude rows are inspectable when PREVIEW_CLAUDE_SOURCES=True, but are NEVER
|
| 722 |
+
included in training unless separately authorised.
|
| 723 |
+
"""
|
| 724 |
+
tok = ns['load_tokenizer']()
|
| 725 |
+
streams = select_streams(ns, for_inspect=True)
|
| 726 |
+
results = {}
|
| 727 |
+
out = Path(P5.WORK_DIR).resolve() / 'phase5_token_audit.json'
|
| 728 |
+
out.parent.mkdir(parents=True, exist_ok=True)
|
| 729 |
+
print(f'π TOKEN AUDIT | {P5.TOKEN_AUDIT_ROWS} accepted samples/source | tokenizer={getattr(tok,"name_or_path","Orion")}', flush=True)
|
| 730 |
+
for name, stream in streams.items():
|
| 731 |
+
lengths, targets, reasons = [], [], []
|
| 732 |
+
for _ in range(P5.TOKEN_AUDIT_ROWS):
|
| 733 |
+
try:
|
| 734 |
+
category, case = stream.next_case(tokenizer=tok)
|
| 735 |
+
if category == 'agent':
|
| 736 |
+
prompt, target, final = case
|
| 737 |
+
p_ids = tok.encode(prompt, add_special_tokens=False)
|
| 738 |
+
t_ids = tok.encode(target, add_special_tokens=False)
|
| 739 |
+
total = len(p_ids) + len(t_ids) + int(bool(final)) + int(tok.bos_token_id is not None)
|
| 740 |
+
target_n = len(t_ids) + int(bool(final))
|
| 741 |
+
reason_n = 0
|
| 742 |
+
else:
|
| 743 |
+
effort = 'Max' if case.get('reason') else 'Instant'
|
| 744 |
+
sample, total = ns['p4_pack'](case, tok, effort, 10**7)
|
| 745 |
+
if sample is None:
|
| 746 |
+
raise ValueError('Sample could not be packed (unusually large).')
|
| 747 |
+
target_n = int((sample[1] != -100).sum().item())
|
| 748 |
+
reason_n = len(tok.encode(case.get('reason', ''), add_special_tokens=False))
|
| 749 |
+
lengths.append(total)
|
| 750 |
+
targets.append(target_n)
|
| 751 |
+
reasons.append(reason_n)
|
| 752 |
+
except Exception as exc:
|
| 753 |
+
print(f'β οΈ {name}: stopped after {len(lengths)} samples: {type(exc).__name__}: {exc}',flush=True)
|
| 754 |
+
break
|
| 755 |
+
if not lengths:
|
| 756 |
+
results[name] = dict(samples=0, error='no accepted samples')
|
| 757 |
+
continue
|
| 758 |
+
ordered = sorted(lengths)
|
| 759 |
+
pct = lambda q: ordered[min(len(ordered)-1, math.ceil(q*len(ordered))-1)]
|
| 760 |
+
result = dict(samples=len(lengths), total_tokens_min=min(lengths),
|
| 761 |
+
total_tokens_median=statistics.median(lengths), total_tokens_p90=pct(.9),
|
| 762 |
+
total_tokens_p99=pct(.99), total_tokens_max=max(lengths),
|
| 763 |
+
target_tokens_median=statistics.median(targets),
|
| 764 |
+
reasoning_tokens_median=statistics.median(reasons),
|
| 765 |
+
above_8k=sum(x>8192 for x in lengths), above_16k=sum(x>16384 for x in lengths),
|
| 766 |
+
above_32k=sum(x>32768 for x in lengths),
|
| 767 |
+
preview_only=stream.spec['rights']=='claude' and not P5.ALLOW_CLAUDE_TEACHER_DATA,
|
| 768 |
+
rejected=dict(stream.rejected))
|
| 769 |
+
results[name] = result
|
| 770 |
+
print(f' {name:26s} n={len(lengths):3d} median={result["total_tokens_median"]:7g} '
|
| 771 |
+
f'p90={result["total_tokens_p90"]:6d} max={result["total_tokens_max"]:6d} '
|
| 772 |
+
f'target_median={result["target_tokens_median"]:6g} '
|
| 773 |
+
f'over32k={result["above_32k"]}',flush=True)
|
| 774 |
+
out.write_text(json.dumps(dict(model=P5.NAME, sample_size=P5.TOKEN_AUDIT_ROWS,
|
| 775 |
+
note='Sample statistics of accepted rows, not full-dataset token totals or training permission.',
|
| 776 |
+
sources=results), indent=2, ensure_ascii=False))
|
| 777 |
+
print(f'β
Saved measured token counts: {out}',flush=True)
|
| 778 |
+
|
| 779 |
+
|
| 780 |
+
def train(ns, smoke=False):
|
| 781 |
+
torch = ns['torch']
|
| 782 |
+
if not torch.cuda.is_available(): raise RuntimeError('Orion Phase5 requires a CUDA GPU.')
|
| 783 |
+
if P5.REQUIRE_HUB_UPLOAD and not os.environ.get('HF_TOKEN'):
|
| 784 |
+
os.environ['HF_TOKEN'] = getpass.getpass('HF write token (hidden): ').strip()
|
| 785 |
+
ns['HF_TOKEN'] = os.environ['HF_TOKEN']
|
| 786 |
+
ns['p3_prepare_gpu']()
|
| 787 |
+
ns['p4_enable_rope']()
|
| 788 |
+
work = Path(P5.WORK_DIR).resolve();work.mkdir(parents=True,exist_ok=True)
|
| 789 |
+
start, folder, mf = pick_checkpoint(ns,work)
|
| 790 |
+
if start >= P5.STEPS:
|
| 791 |
+
print(f'π Phase5 already finished: {start}/{P5.STEPS}');return
|
| 792 |
+
parent = mf.get('phase5_parent') if start else (
|
| 793 |
+
'checkpoints/phase4-step-'+str(mf['step']).zfill(6))
|
| 794 |
+
arch = ns['p4_make_arch'](mf['architecture'])
|
| 795 |
+
ns['select_attention'](arch)
|
| 796 |
+
model = ns['build_model'](arch)
|
| 797 |
+
ns['restore_model_shards'](model,folder,mf['model_shards'])
|
| 798 |
+
model.set_specialization_scale(1.0);model.train()
|
| 799 |
+
tok = ns['load_tokenizer']()
|
| 800 |
+
optimizer = torch.optim.AdamW(model.parameters(), lr=P5.LEARNING_RATE,
|
| 801 |
+
betas=(.9,.95),eps=1e-8,weight_decay=.01,fused=True)
|
| 802 |
+
if start and (folder/'optimizer.pt').exists():
|
| 803 |
+
optimizer.load_state_dict(torch.load(folder/'optimizer.pt',map_location=ns['DEVICE'],weights_only=True))
|
| 804 |
+
elif start:
|
| 805 |
+
print('β οΈ AdamW moments reset on resume (SAVE_OPTIMIZER=False).')
|
| 806 |
+
streams = select_streams(ns, mf.get('phase5_source_positions',{}) if start else {})
|
| 807 |
+
rng = random.Random(P5.SEED + start*100003)
|
| 808 |
+
print(f'π PHASE5 {start}/{P5.STEPS} -> {P5.NAME} | sources={list(streams)}',flush=True)
|
| 809 |
+
if smoke: print('π§ͺ SMOKE MODE: one update, no checkpoint upload.',flush=True)
|
| 810 |
+
for step in range(start+1, (start+2 if smoke else P5.STEPS+1)):
|
| 811 |
+
cap = cap_for_step(step)
|
| 812 |
+
t0 = time.monotonic()
|
| 813 |
+
while True:
|
| 814 |
+
optimizer.zero_grad(set_to_none=True)
|
| 815 |
+
counts = collections.Counter()
|
| 816 |
+
total_ce = 0.
|
| 817 |
+
longest = 0
|
| 818 |
+
accum = 1 if smoke else P5.GRAD_ACCUM
|
| 819 |
+
try:
|
| 820 |
+
for _ in range(accum):
|
| 821 |
+
ns['ACTIVE_CONTEXT_LIMIT'] = cap
|
| 822 |
+
label,(x,y) = choose_sample(ns,streams,tok,cap,rng)
|
| 823 |
+
longest = max(longest,len(x))
|
| 824 |
+
with torch.autocast(device_type='cuda',dtype=torch.bfloat16,cache_enabled=True):
|
| 825 |
+
ce, aux = model(x.unsqueeze(0).to(ns['DEVICE']), y.unsqueeze(0).to(ns['DEVICE']))
|
| 826 |
+
loss = (ce + .25*aux)/accum
|
| 827 |
+
if not torch.isfinite(loss).item(): raise RuntimeError('Nonfinite Phase5 loss')
|
| 828 |
+
loss.backward()
|
| 829 |
+
counts[label] += 1
|
| 830 |
+
total_ce += float(ce.detach())
|
| 831 |
+
break
|
| 832 |
+
except torch.cuda.OutOfMemoryError:
|
| 833 |
+
optimizer.zero_grad(set_to_none=True)
|
| 834 |
+
torch.cuda.empty_cache()
|
| 835 |
+
if cap <= 8192: raise
|
| 836 |
+
cap //= 2
|
| 837 |
+
print(f'β οΈ OOM: retrying with whole samples <= {cap} tokens.',flush=True)
|
| 838 |
+
grad = torch.nn.utils.clip_grad_norm_(model.parameters(),1.0)
|
| 839 |
+
if not torch.isfinite(grad).item(): raise RuntimeError('Nonfinite gradient norm')
|
| 840 |
+
lr = ns['lr_at'](step,P5.STEPS,P5.LEARNING_RATE,warmup=50)
|
| 841 |
+
for group in optimizer.param_groups: group['lr'] = lr
|
| 842 |
+
optimizer.step()
|
| 843 |
+
if smoke or step == start+1 or step%P5.LOG_EVERY == 0:
|
| 844 |
+
print(f'π₯ PHASE5 {step}/{P5.STEPS} CE={total_ce/accum:.4f} '
|
| 845 |
+
f'grad={float(grad):.3f} lr={lr:.2e} cap={cap} '
|
| 846 |
+
f'longest={longest} elapsed={time.monotonic()-t0:.1f}s sources={dict(counts)}',flush=True)
|
| 847 |
+
if smoke:
|
| 848 |
+
print('β
Smoke complete. No checkpoint uploaded.');break
|
| 849 |
+
if step%P5.EVAL_EVERY==0:
|
| 850 |
+
val = eval_loss(ns,model,tok)
|
| 851 |
+
print(f'π Fixed diagnostic CE={val:.4f} (not a reasoning accuracy score)',flush=True)
|
| 852 |
+
if step%P5.SAVE_EVERY==0 or step==P5.STEPS:
|
| 853 |
+
save(ns,model,tok,optimizer,arch,step,streams,work,parent)
|
| 854 |
+
print('π Orion Phase5 run complete or smoke finished.',flush=True)
|
| 855 |
+
|
| 856 |
+
|
| 857 |
+
def rescue(ns):
|
| 858 |
+
work=Path(P5.WORK_DIR).resolve()
|
| 859 |
+
valid=[]
|
| 860 |
+
for path in work.glob('phase5-step-*'):
|
| 861 |
+
try:
|
| 862 |
+
mf=ns['check_checkpoint_folder'](path,('orion-phase5-1',))
|
| 863 |
+
valid.append((int(mf['step']),path))
|
| 864 |
+
except (ValueError,OSError,KeyError): pass
|
| 865 |
+
if not valid: raise RuntimeError('No complete local Phase5 checkpoint to rescue.')
|
| 866 |
+
step, path=max(valid)
|
| 867 |
+
ns['upload_checkpoint'](path,P5.UPLOAD_REPO)
|
| 868 |
+
ns['combo_uploaded_files'](path,P5.UPLOAD_REPO)
|
| 869 |
+
print('β
Rescued verified Phase5 step',step)
|
| 870 |
+
|
| 871 |
+
|
| 872 |
+
def main():
|
| 873 |
+
print(f'π€ ORION PHASE5 | mode={P5.MODE} | {P5.NAME}')
|
| 874 |
+
if P5.MODE=='selftest': return selftest()
|
| 875 |
+
if P5.MODE not in ('inspect','token_audit','smoke','train','rescue'):
|
| 876 |
+
raise ValueError('MODE must be selftest, inspect, token_audit, smoke, train, or rescue')
|
| 877 |
+
if P5.MODE in ('train','smoke','rescue') and P5.REQUIRE_HUB_UPLOAD and not os.environ.get('HF_TOKEN'):
|
| 878 |
+
os.environ['HF_TOKEN'] = getpass.getpass('Hugging Face write token (hidden): ').strip()
|
| 879 |
+
ns=load_orion_runtime()
|
| 880 |
+
if P5.MODE=='inspect': return inspect(ns)
|
| 881 |
+
if P5.MODE=='token_audit': return token_audit(ns)
|
| 882 |
+
if P5.MODE=='smoke': return train(ns,smoke=True)
|
| 883 |
+
if P5.MODE=='train': return train(ns,smoke=False)
|
| 884 |
+
if P5.MODE=='rescue': return rescue(ns)
|
| 885 |
+
|
| 886 |
+
|
| 887 |
+
if __name__=='__main__': main()
|
| 888 |
+
|
| 889 |
+
|
| 890 |
+
return
|
| 891 |
+
|
| 892 |
+
|
| 893 |
+
if __name__ == "__main__":
|
| 894 |
+
app.run()
|
checkpoints/phase5-step-004900/orion_unified_trainer.py
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
checkpoints/phase5-step-004900/tokenizer/tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
checkpoints/phase5-step-004900/tokenizer/tokenizer_config.json
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_prefix_space": true,
|
| 3 |
+
"backend": "tokenizers",
|
| 4 |
+
"bos_token": "<s>",
|
| 5 |
+
"clean_up_tokenization_spaces": false,
|
| 6 |
+
"eos_token": "</s>",
|
| 7 |
+
"is_local": false,
|
| 8 |
+
"legacy": false,
|
| 9 |
+
"local_files_only": false,
|
| 10 |
+
"model_max_length": 1000000000000000019884624838656,
|
| 11 |
+
"pad_token": null,
|
| 12 |
+
"sp_model_kwargs": {},
|
| 13 |
+
"spaces_between_special_tokens": false,
|
| 14 |
+
"tokenizer_class": "TokenizersBackend",
|
| 15 |
+
"unk_token": "<unk>",
|
| 16 |
+
"use_default_system_prompt": false
|
| 17 |
+
}
|