Update DPO family dataset: basic_arithmetic_qa

README.md

@@ -6,20 +6,26 @@ configs:
     path: data/*.jsonl
 ---
 
-# SLM Synthetic DPO
+# SLM Synthetic DPO — basic_arithmetic_qa
 
 ## Summary
 
-Synthetic preference-pair dataset containing chosen and rejected assistant responses.
+Synthetic preference-pair dataset containing chosen and rejected assistant responses for `basic_arithmetic_qa`.
 
 ## 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`
+- Total pairs: `6429`
+- Signal: `basic_arithmetic_qa`
 - Language: English
 
+## Signal Metadata
+
+- Category: `direct_arithmetic`
+- Difficulty: `1`
+- Template family: `direct_addition`
+- Eval family: `basic_arithmetic_qa`
+
 ## Schema
 
 ```json
@@ -46,7 +52,7 @@ Synthetic preference-pair dataset containing chosen and rejected assistant respo
 
 ## Intended Use
 
-Use this dataset for DPO-style preference optimization experiments.
+Use this dataset for DPO-style preference optimization experiments targeting this signal.
 
 ## Limitations