Upload README.md with huggingface_hub

README.md

@@ -1,62 +1,71 @@
 ---
 license: other
-task_categories:
-- text-generation
 language:
 - en
+- zh
 tags:
 - sft
-- agentic
 - coding
+- agentic
 - tool-use
-- reasoning
+- math
+- stem
+- grounded-generation
 - anti-hallucination
 ---
 
-# Agentic Dataset v3
+# Agentic Dataset v6 Dynamic
 
-100K-row normalized SFT corpus for agentic/coding/tool-use fine-tuning.
+Dynamically allocated SFT dataset.
 
-## Domain mix
+Requested target:
+100,000
 
-- coding: 33,040
-- agent: 25,960
-- tool: 17,611
-- math: 16,520
-- instruction: 1,776
-- stem: 443
+Actual target:
+100,000
 
-## Quality
+## Domain distribution
 
-- Anti-hallucination / grounding flagged: 24,584
-- Tool-call rows: 11,887
-- Tool-grounded rows: 23,625
-- Coding rows with testing signal: 2,235
-- Math signal: 16,520
-- Temporal unverified rows: 3,433
-- Fake external-action claims: 0
+coding: 32,200
+agent: 15,235
+tool: 22,565
+math: 14,000
+instruction: 10,000
+stem: 6,000
 
-## Length
+## Reliability
+
+anti_hallucination:
+30,512
+
+tool_call_rows:
+22,265
 
-- P50 chars: 1,730
-- P95 chars: 7,461
-- P99 chars: 15,532
-- P95 approximate tokens: 1,865
+tool_result_rows:
+21,972
 
-## Schema
+tool_grounded_rows:
+21,971
+
+verified_rows:
+18,384
+
+uncertainty_rows:
+1,058
+
+fake_tool_claims:
+0
+
+## Length
 
-Each row contains:
+P50 chars:
+2,038
 
-- `id`
-- `messages`
-- `tools`
-- `domain`
-- `source`
-- `teacher_model`
-- `quality_score`
-- `anti_hallucination`
-- `quality_flags`
+P95 chars:
+14,241
 
-## Important
+P99 chars:
+26,783
 
-This dataset shapes behavior toward grounded responses; it does not mathematically guarantee zero hallucinations.
+The anti-hallucination flag is a training-data quality signal,
+not a guarantee of zero hallucinations at inference time.