File size: 11,551 Bytes
df0b00f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
{
  "data_designer": {
    "columns": [
      {
        "name": "example_id",
        "drop": false,
        "allow_resize": false,
        "column_type": "sampler",
        "sampler_type": "uuid",
        "params": {
          "prefix": null,
          "short_form": false,
          "uppercase": false,
          "sampler_type": "uuid"
        },
        "conditional_params": {},
        "convert_to": null
      },
      {
        "name": "batch_id",
        "drop": false,
        "allow_resize": false,
        "column_type": "sampler",
        "sampler_type": "uuid",
        "params": {
          "prefix": null,
          "short_form": false,
          "uppercase": false,
          "sampler_type": "uuid"
        },
        "conditional_params": {},
        "convert_to": null
      },
      {
        "name": "source_id",
        "drop": false,
        "allow_resize": false,
        "column_type": "expression",
        "expr": "\"tatsu-lab/alpaca\"                                                                                        ",
        "dtype": "str"
      },
      {
        "name": "language",
        "drop": false,
        "allow_resize": false,
        "column_type": "expression",
        "expr": "\"en\"                                                                                                                                                                              ",
        "dtype": "str"
      },
      {
        "name": "pipeline_version",
        "drop": false,
        "allow_resize": false,
        "column_type": "expression",
        "expr": "\"recipe_v0.1\"",
        "dtype": "str"
      },
      {
        "name": "ai_evaluation",
        "drop": false,
        "allow_resize": false,
        "column_type": "llm-structured",
        "prompt": "You are a rigorous dataset-selection judge operating under a methodological data curation framework.\n\nYour task is to evaluate ONE dataset example for its suitability as a training example based on methodological value, evidence discipline, structural completeness, correctness, and traceability.\n\nIMPORTANT:\n\n* Evaluate the example ONLY for selection quality at this stage.\n* DO NOT assign any curriculum capability level such as L0-L13.\n* DO NOT infer or invent a level.\n* DO NOT treat mathematical difficulty, domain difficulty, answer length, reasoning length, or the presence of step-by-step calculations as sufficient evidence of methodological value.\n* Your judgment is a recommendation for downstream governance checks. It is NOT the final ACCEPT / REVIEW / REJECT decision.\n* A quality_score of 4 or 5 does NOT automatically mean ACCEPT.\n* Do not fabricate missing information.\n* If important information is missing or the decision is uncertain, reflect that uncertainty through confidence and reason.\n\nINPUT\n\nInstruction: {{ instruction }}\nInput Context: {{ input }}\nGenerated Output: {{ output }}\n\nEVALUATION OBJECTIVE\n\nDetermine whether this example contains useful, reliable, traceable training value and whether it should be considered a candidate for downstream selection checks.\n\nEvaluate the following dimensions:\n\n1. METHODOLOGICAL VALUE\n   Assess whether the example demonstrates a useful problem-solving, reasoning, verification, evidence, instruction-following, or other methodological behavior.\n\nDo NOT give a high score merely because:\n\n* the arithmetic is correct,\n* the answer is fluent,\n* the solution is long,\n* the solution is step-by-step,\n* the problem is difficult,\n* or the formatting looks polished.\n\nThe example should contain meaningful behavior that can contribute to disciplined model training.\n\n2. CORRECTNESS AND FACT PRESERVATION\n   Check whether:\n\n* the answer is consistent with the question,\n* calculations or logical operations are correct,\n* important facts from the question are preserved,\n* the reasoning does not contradict the input,\n* and the final answer follows from the presented reasoning.\n\nIdentify substantive errors rather than stylistic preferences.\n\n3. EVIDENCE AND TRACEABILITY\n   Check whether the reasoning can be verified directly from the provided question and answer.\n\nPrefer examples where:\n\n* the necessary information is present,\n* reasoning steps can be traced to the input,\n* important claims are supported by the available context,\n* and the answer does not rely on invented external facts.\n\nDo not treat an explanation written by the answer itself as independent evidence. The evidence must be traceable to the provided example.\n\n4. COMPLETENESS AND ATOMICITY\n   Check whether the example is sufficiently complete to be useful as a standalone training example.\n\nIdentify:\n\n* missing reasoning required to justify the answer,\n* abrupt truncation,\n* incomplete answers,\n* corrupted content,\n* unresolved references,\n* missing context,\n* or evidence that has been cut.\n\nDo not reject an example merely because it is concise. Concision is acceptable when the example remains complete and methodologically useful.\n\n5. HALLUCINATION, OVERREACH, AND UNSUPPORTED INFERENCE\n   Check whether the answer:\n\n* introduces facts not supported by the question,\n* makes unsupported assumptions,\n* performs unjustified inference,\n* changes facts from the input,\n* or presents uncertain information as certain.\n\nThese are important quality problems.\n\n6. DUPLICATION / LOW-VALUE BEHAVIOR\n   Consider whether the example appears to provide little methodological value because it is trivial, redundant, malformed, or contains no useful behavior beyond a bare answer.\n\nDo not reject solely because the task is simple. A simple example can still have methodological value.\n\n7. OBJECTIVE BLOCKERS\n   Identify concrete blockers such as:\n\n* incorrect answer,\n* contradictory reasoning,\n* corrupted or incomplete text,\n* missing essential context,\n* evidence-cut,\n* untraceable claims,\n* or severe formatting/content problems.\n\nDo not invent blockers.\n\nQUALITY SCORE\n\nAssign an integer quality_score from 1 to 5.\n\nUse this rubric:\n\n5 = Excellent candidate.\nThe example is correct, complete, traceable, methodologically useful, and has no significant blocker.\n\n4 = Strong candidate.\nThe example is substantially correct and useful, with only minor limitations that do not seriously reduce its training value.\n\n3 = Borderline / requires review.\nThe example has some useful value but also has meaningful ambiguity, incompleteness, weakness, or uncertainty that should be reviewed.\n\n2 = Poor candidate.\nThe example contains a significant quality problem, methodological weakness, unsupported inference, incorrect reasoning, or other issue that substantially reduces its training value.\n\n1 = Clearly unsuitable.\nThe example is corrupted, fundamentally incorrect, severely incomplete, untraceable, or otherwise has no reliable training value.\n\nIMPORTANT:\nThe score is NOT a final governance decision.\n\nSELECTION CANDIDATE\n\nSet selection_candidate to true only when the example appears to be a strong downstream candidate based on the criteria above.\n\nSet selection_candidate to false when there is a clear quality problem.\n\nIf the example is genuinely ambiguous or requires human/governance review, do NOT hide that uncertainty. Use a lower confidence and explain the ambiguity.\n\nselection_candidate is only an AI recommendation.\nIt does NOT mean final ACCEPT.\nThe downstream Checks and governance rules determine ACCEPT / REVIEW / REJECT.\n\nCONFIDENCE\n\nSet confidence to a number between 0.0 and 1.0.\n\nUse confidence to represent how certain you are about your evaluation.\n\nHigh confidence:\n\n* evidence is clear,\n* correctness is directly verifiable,\n* and the quality judgment is unambiguous.\n\nLower confidence:\n\n* important context is missing,\n* quality is borderline,\n* methodological value is ambiguous,\n* evidence is insufficient,\n* or multiple interpretations are plausible.\n\nDo not use confidence merely to express how difficult the math is.\n\nEVIDENCE\n\nProvide a concise direct quote or exact snippet from the provided Question and/or Answer that supports the evaluation.\n\nThe evidence must be traceable to the supplied input.\n\nDo NOT invent evidence.\nDo NOT use external knowledge as evidence.\nDo NOT provide evidence that is not present in the input.\n\nREASON\n\nProvide a concise explanation of WHY the example received its score and recommendation.\n\nThe reason must focus on methodological quality and concrete evidence.\n\nDo not use generic praise such as:\n\n* \"This is a good example.\"\n* \"The reasoning is pedagogical.\"\n* \"The answer is well written.\"\n\nInstead identify the specific behavior, correctness, evidence, limitation, or blocker that justifies the judgment.\n\nFINAL OUTPUT\n\nReturn ONLY a structured object containing exactly these fields:\n\n{\n\"quality_score\": integer,\n\"selection_candidate\": boolean,\n\"reason\": \"concise evidence-based explanation\",\n\"confidence\": float,\n\"evidence\": \"direct quote or exact traceable snippet\"\n}\n\nFINAL RULES\n\n* Never assign L0-L13.\n* Never invent missing facts.\n* Never fabricate evidence.\n* Never treat score >= 4 as automatic acceptance.\n* Never collapse uncertainty into false certainty.\n* Never use domain difficulty as a substitute for methodological value.\n* Never use answer length as a substitute for quality.\n* Never use step-by-step formatting alone as a substitute for methodological value.\n* Preserve the distinction between AI recommendation and final governance decision.\n* When uncertain, state the uncertainty through the reason and confidence.\n",
        "model_alias": "selection_judge",
        "system_prompt": null,
        "multi_modal_context": null,
        "tool_alias": null,
        "with_trace": "none",
        "extract_reasoning_content": false,
        "output_format": {
          "quality_score": "number",
          "selection_candidate": "boolean",
          "reason": "string",
          "confidence": "number",
          "evidence": "string"
        }
      },
      {
        "name": "instruction",
        "drop": false,
        "allow_resize": false,
        "column_type": "seed-dataset"
      },
      {
        "name": "input",
        "drop": false,
        "allow_resize": false,
        "column_type": "seed-dataset"
      },
      {
        "name": "output",
        "drop": false,
        "allow_resize": false,
        "column_type": "seed-dataset"
      },
      {
        "name": "text",
        "drop": false,
        "allow_resize": false,
        "column_type": "seed-dataset"
      }
    ],
    "model_configs": [
      {
        "alias": "selection_judge",
        "model": "gemini-3.5-flash-lite",
        "inference_parameters": {
          "generation_type": "chat-completion",
          "max_parallel_requests": 1,
          "timeout": null,
          "extra_body": null,
          "temperature": 0.7,
          "top_p": null,
          "max_tokens": null
        },
        "provider": "Gemini_Flash_Free",
        "skip_health_check": false
      }
    ],
    "tool_configs": [],
    "seed_config": {
      "source": {
        "seed_type": "hf",
        "path": "datasets/tatsu-lab/alpaca/data/**/*.parquet",
        "token": null,
        "endpoint": "https://huggingface.co"
      },
      "sampling_strategy": "ordered",
      "selection_strategy": null
    },
    "constraints": null,
    "profilers": null,
    "processors": null
  },
  "library_version": "0.5.4"
}