Update DPO family dataset: code_generation_function

README.md

@@ -0,0 +1,53 @@
+---
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/*.jsonl
+---
+
+# SLM Synthetic DPO
+
+## Summary
+
+Synthetic preference-pair dataset containing chosen and rejected assistant responses.
+
+## Dataset
+
+- Dataset type: `preference optimization`
+- Total pairs: `90000`
+- Signals: `ai_concept_explanation, basic_arithmetic_qa, capital_city_qa, clear_sky_color_qa, code_explanation_no_code, code_expression_result, code_generation_function, direct_division, direct_subtraction, function_completion_body_only, list_exact_n_items, private_or_unverifiable_company_fact, repeat_exact_n_times, short_factual_stop_behavior`
+- Teacher model: `deepseek/deepseek-v4-flash`
+- Language: English
+
+## Schema
+
+```json
+{
+  "id": "string",
+  "prompt": [
+    {"role": "user", "content": "string"}
+  ],
+  "chosen": [
+    {"role": "assistant", "content": "string"}
+  ],
+  "rejected": [
+    {"role": "assistant", "content": "string"}
+  ],
+  "metadata": {
+    "category": "string",
+    "difficulty": 1,
+    "template_family": "string",
+    "eval_family": "string",
+    "failure_mode": "string"
+  }
+}
+```
+
+## Intended Use
+
+Use this dataset for DPO-style preference optimization experiments.
+
+## Limitations
+
+This dataset is synthetic and encodes targeted preference behavior. Inspect chosen/rejected pairs before training.

artifacts/coverage.json

@@ -0,0 +1,74 @@
+{
+  "dataset_type": "dpo",
+  "row_count": 90000,
+  "files": {
+    "data/dpo/runs/dpo-target-001/datasets/ai_concept_explanation.jsonl": 6429,
+    "data/dpo/runs/dpo-target-001/datasets/basic_arithmetic_qa.jsonl": 6429,
+    "data/dpo/runs/dpo-target-001/datasets/capital_city_qa.jsonl": 6429,
+    "data/dpo/runs/dpo-target-001/datasets/clear_sky_color_qa.jsonl": 6429,
+    "data/dpo/runs/dpo-target-001/datasets/code_explanation_no_code.jsonl": 6429,
+    "data/dpo/runs/dpo-target-001/datasets/code_expression_result.jsonl": 6429,
+    "data/dpo/runs/dpo-target-001/datasets/code_generation_function.jsonl": 6429,
+    "data/dpo/runs/dpo-target-001/datasets/direct_division.jsonl": 6429,
+    "data/dpo/runs/dpo-target-001/datasets/direct_subtraction.jsonl": 6428,
+    "data/dpo/runs/dpo-target-001/datasets/function_completion_body_only.jsonl": 6428,
+    "data/dpo/runs/dpo-target-001/datasets/list_exact_n_items.jsonl": 6428,
+    "data/dpo/runs/dpo-target-001/datasets/private_or_unverifiable_company_fact.jsonl": 6428,
+    "data/dpo/runs/dpo-target-001/datasets/repeat_exact_n_times.jsonl": 6428,
+    "data/dpo/runs/dpo-target-001/datasets/short_factual_stop_behavior.jsonl": 6428
+  },
+  "categories": {
+    "code_expression_evaluation": 6429,
+    "code_generation": 12857,
+    "concise_factual_qa": 12858,
+    "controlled_verbosity": 6428,
+    "direct_arithmetic": 19286,
+    "exact_output_format_control": 12856,
+    "general_instruction_following": 12858,
+    "private_info_restraint": 6428
+  },
+  "eval_families": {
+    "ai_concept_explanation": 6429,
+    "basic_arithmetic_qa": 6429,
+    "capital_city_qa": 6429,
+    "clear_sky_color_qa": 6429,
+    "code_explanation_no_code": 6429,
+    "code_expression_result": 6429,
+    "code_generation_function": 6429,
+    "direct_division": 6429,
+    "direct_subtraction": 6428,
+    "function_completion_body_only": 6428,
+    "list_exact_n_items": 6428,
+    "private_or_unverifiable_company_fact": 6428,
+    "repeat_exact_n_times": 6428,
+    "short_factual_stop_behavior": 6428
+  },
+  "template_families": {
+    "capital_city_direct": 6429,
+    "code_explanation_plain_text": 6429,
+    "common_fact_color": 6429,
+    "direct_addition": 6429,
+    "direct_division": 6429,
+    "direct_subtraction": 6428,
+    "list_exact_count": 6428,
+    "private_company_metric": 6428,
+    "python_expression_result": 6429,
+    "python_function_body_only": 6428,
+    "python_function_code_only": 6429,
+    "repeat_word_count": 6428,
+    "short_ai_definition": 6429,
+    "short_factual_answer": 6428
+  },
+  "difficulty_counts": {
+    "1": 51428,
+    "2": 38572
+  },
+  "failure_modes": {
+    "code_includes_explanation": 19286,
+    "format_violation": 12856,
+    "unknown_fact_fabrication": 6428,
+    "verbosity_mismatch": 6428,
+    "wrong_factual_answer": 19287,
+    "wrong_numeric_answer": 25715
+  }
+}