Add A10 long-memory and exact-answer specialist data

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+train.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,47 @@
+---
+pretty_name: chatalpaca-multiturn-enriched-3
+task_categories:
+- text-generation
+language:
+- en
+---
+
+# chatalpaca-multiturn-enriched-3
+
+This dataset combines the existing Samantha A10 multiturn corpus with new long-memory and exact-answer specialist conversations.
+
+## Splits
+
+- `train`: 18,914 rows (existing, manual-evaluation, and generated rows)
+- No separate validation split is published; all records remain in `train`.
+- Total: 18,914 rows
+
+## Composition
+
+- Existing source artifact: `BRlkl/chatalpaca-multiturn-enriched-2.1`
+- New long-memory rows: 8,000
+- New arithmetic rows: 1,000
+- New logic rows: 1,000
+- New conversational-QA rows: 1,000
+- New exact-transformation rows: 500
+
+The long-memory source seeds come from `HuggingFaceH4/no_robots`. Specialist grounding comes from `openai/gsm8k`, `tasksource/proofwriter`, and `stanfordnlp/coqa`. Exact transformations are generated from deterministic specifications.
+
+## Generation contract
+
+- Local vLLM writer model: `Qwen/Qwen3.6-27B`
+- vLLM server is started and stopped by the generation runner
+- All pending API requests are submitted immediately; vLLM owns the queue
+- No LLM judge pass
+- Long-memory rows use three stages: source expansion, memory revision, and family specialization
+- Arithmetic, logic, QA, and transformation rows use a separate two-stage exact-answer pipeline
+- Every new row contains exactly 12 user/assistant pairs (24 messages)
+- Message roles use `user_message_N` and `assistant_message_N`; existing rows are converted to the same schema during merge
+- Each row receives one qualitative message-length profile: `short`, `medium`, `long`, or `varied`
+- No per-message length targets
+- No tokenizer-based conversation-length cap
+- No automatic regex quality verifier
+- JSON shape, indexed-role order, and exact specialist final answers are checked deterministically
+- The counterfactual same-surface family is one conversation with two states and different context-conditioned answers
+
+The raw generation manifest and validation report are included in the repository.

manifest.json

@@ -0,0 +1,165 @@
+{
+  "created": "2026-08-08",
+  "models": [
+    "Qwen/Qwen3.6-27B"
+  ],
+  "judge_enabled": false,
+  "existing_dataset_repo_id": "BRlkl/chatalpaca-multiturn-enriched-2.1",
+  "hf_dataset_repo_id": "BRlkl/chatalpaca-multiturn-enriched-3",
+  "source_datasets": {
+    "long_memory": "HuggingFaceH4/no_robots",
+    "arithmetic": "openai/gsm8k",
+    "logic": "tasksource/proofwriter",
+    "qa": "stanfordnlp/coqa",
+    "transformation": "deterministic_exact_transformations"
+  },
+  "requested_new_rows": 11500,
+  "api_tasks": 11500,
+  "generation_pipeline": {
+    "long_memory": "three local-vLLM stages: expansion, memory revision, family specialization",
+    "specialists": "two local-vLLM stages: exact-answer draft, memory revision"
+  },
+  "pair_count_contract": "exactly 12 user/assistant pairs for every new row",
+  "message_role_contract": "user_message_N / assistant_message_N",
+  "length_profile_contract": "one coarse row-level profile; no per-message length enforcement",
+  "automatic_regex_verifiers": false,
+  "tokenizer_or_conversation_token_cap": false,
+  "scenario_row_counts": {
+    "clarify_then_resolve": 267,
+    "coqa_contextual_short_answer": 1000,
+    "coreference_disambiguation": 267,
+    "counterfactual_option_swap": 267,
+    "counterfactual_same_surface_twins": 800,
+    "deferred_question_then_answer": 267,
+    "distributed_multi_turn_synthesis": 800,
+    "early_fact_late_recall": 267,
+    "exact_arithmetic_multiturn": 1000,
+    "exact_logic_multiturn": 1000,
+    "exact_transformation_chronological_order": 83,
+    "exact_transformation_deduplicate": 83,
+    "exact_transformation_extract_field": 83,
+    "exact_transformation_reverse_string": 83,
+    "exact_transformation_select_index": 84,
+    "exact_transformation_sort_items": 84,
+    "hard_topic_reset": 267,
+    "interrupting_followup_chain": 267,
+    "late_constraint_reversal": 267,
+    "late_request_for_summary_of_prior_state": 267,
+    "long_explanation_then_specific_followup": 267,
+    "preserve_opening_retarget_later": 267,
+    "reference_previous_assistant_statement": 267,
+    "rolling_state_supersession": 800,
+    "same_surface_final_question_new_context": 266,
+    "self_correction_memory": 266,
+    "stacked_entities_same_type": 266,
+    "topic_pivot_then_return": 266,
+    "two_thread_tracking": 266,
+    "user_attention_test": 266,
+    "very_early_anchor_very_late_recall": 800
+  },
+  "length_profile_row_counts": {
+    "long": 2949,
+    "medium": 2311,
+    "short": 2836,
+    "varied": 3404
+  },
+  "specialist_row_counts": {
+    "arithmetic": 1000,
+    "logic": 1000,
+    "qa": 1000,
+    "transformation": 500
+  },
+  "global_concurrency": "unbounded_client_submission",
+  "temperature": 0.7,
+  "top_p": 0.8,
+  "max_retries": 2,
+  "http_retries": 3,
+  "generation_rounds": 3,
+  "transport": "local_vllm",
+  "request_submission": "all pending API tasks scheduled at once; no client semaphore or worker pool",
+  "client_max_connections": null,
+  "vllm": {
+    "managed_server": true,
+    "model": "Qwen/Qwen3.6-27B",
+    "endpoint": "http://127.0.0.1:8000/v1",
+    "serve_options": {
+      "tensor_parallel_size": 2,
+      "dtype": "bfloat16",
+      "gpu_memory_utilization": 0.95,
+      "max_model_len": 32768,
+      "max_num_seqs": 1024,
+      "max_num_batched_tokens": 65536,
+      "enable_chunked_prefill": true,
+      "enable_prefix_caching": true,
+      "language_model_only": true,
+      "reasoning_parser": "qwen3",
+      "default_chat_template_kwargs": "{\"enable_thinking\": false}"
+    }
+  },
+  "published_counts": {
+    "existing_rows": 7944,
+    "new_rows": 10970,
+    "train_rows": 18914,
+    "validation_rows": 0,
+    "total_rows": 18914,
+    "generation_complete": false
+  },
+  "generated_validation": {
+    "rows": 10970,
+    "source_kind_counts": {
+      "generated_long_memory": 7556,
+      "specialist_arithmetic": 998,
+      "specialist_logic": 976,
+      "specialist_qa": 944,
+      "specialist_transformation": 496
+    },
+    "scenario_counts": {
+      "clarify_then_resolve": 254,
+      "coqa_contextual_short_answer": 944,
+      "coreference_disambiguation": 254,
+      "counterfactual_option_swap": 251,
+      "counterfactual_same_surface_twins": 755,
+      "deferred_question_then_answer": 250,
+      "distributed_multi_turn_synthesis": 764,
+      "early_fact_late_recall": 253,
+      "exact_arithmetic_multiturn": 998,
+      "exact_logic_multiturn": 976,
+      "exact_transformation_chronological_order": 83,
+      "exact_transformation_deduplicate": 82,
+      "exact_transformation_extract_field": 83,
+      "exact_transformation_reverse_string": 80,
+      "exact_transformation_select_index": 84,
+      "exact_transformation_sort_items": 84,
+      "hard_topic_reset": 251,
+      "interrupting_followup_chain": 256,
+      "late_constraint_reversal": 253,
+      "late_request_for_summary_of_prior_state": 255,
+      "long_explanation_then_specific_followup": 256,
+      "preserve_opening_retarget_later": 251,
+      "reference_previous_assistant_statement": 247,
+      "rolling_state_supersession": 753,
+      "same_surface_final_question_new_context": 250,
+      "self_correction_memory": 243,
+      "stacked_entities_same_type": 245,
+      "topic_pivot_then_return": 250,
+      "two_thread_tracking": 254,
+      "user_attention_test": 249,
+      "very_early_anchor_very_late_recall": 762
+    },
+    "length_profile_counts": {
+      "long": 2810,
+      "medium": 2223,
+      "short": 2586,
+      "varied": 3351
+    },
+    "pair_counts": {
+      "12": 10970
+    },
+    "schema_validation": "JSON object, 24 non-empty indexed-role messages, exact specialist answer",
+    "regex_validation": false,
+    "conversation_token_cap": null,
+    "complete": false,
+    "expected_rows": 11500,
+    "missing_rows": 530
+  }
+}

train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb4042890e7f22cdf216b004cab5a3856c97d79f09bfca97e7f8b55574e0a658
+size 86546298

validation_report.md

@@ -0,0 +1,62 @@
+# Generation validation report
+
+```json
+{
+  "rows": 10970,
+  "source_kind_counts": {
+    "generated_long_memory": 7556,
+    "specialist_arithmetic": 998,
+    "specialist_logic": 976,
+    "specialist_qa": 944,
+    "specialist_transformation": 496
+  },
+  "scenario_counts": {
+    "clarify_then_resolve": 254,
+    "coqa_contextual_short_answer": 944,
+    "coreference_disambiguation": 254,
+    "counterfactual_option_swap": 251,
+    "counterfactual_same_surface_twins": 755,
+    "deferred_question_then_answer": 250,
+    "distributed_multi_turn_synthesis": 764,
+    "early_fact_late_recall": 253,
+    "exact_arithmetic_multiturn": 998,
+    "exact_logic_multiturn": 976,
+    "exact_transformation_chronological_order": 83,
+    "exact_transformation_deduplicate": 82,
+    "exact_transformation_extract_field": 83,
+    "exact_transformation_reverse_string": 80,
+    "exact_transformation_select_index": 84,
+    "exact_transformation_sort_items": 84,
+    "hard_topic_reset": 251,
+    "interrupting_followup_chain": 256,
+    "late_constraint_reversal": 253,
+    "late_request_for_summary_of_prior_state": 255,
+    "long_explanation_then_specific_followup": 256,
+    "preserve_opening_retarget_later": 251,
+    "reference_previous_assistant_statement": 247,
+    "rolling_state_supersession": 753,
+    "same_surface_final_question_new_context": 250,
+    "self_correction_memory": 243,
+    "stacked_entities_same_type": 245,
+    "topic_pivot_then_return": 250,
+    "two_thread_tracking": 254,
+    "user_attention_test": 249,
+    "very_early_anchor_very_late_recall": 762
+  },
+  "length_profile_counts": {
+    "long": 2810,
+    "medium": 2223,
+    "short": 2586,
+    "varied": 3351
+  },
+  "pair_counts": {
+    "12": 10970
+  },
+  "schema_validation": "JSON object, 24 non-empty indexed-role messages, exact specialist answer",
+  "regex_validation": false,
+  "conversation_token_cap": null,
+  "complete": false,
+  "expected_rows": 11500,
+  "missing_rows": 530
+}
+```