vm2825/test_beam
BEAM Compact Gemini Scaling Sample This is a wholly synthetic, machine-generated scaling sample inspired by the schema of Mohammadta/BEAM. It is not an official BEAM release and has not received BEAM's human validation. Contents Conversations: 50 Probing questions: 1000 Questions per conversation: 20 Model: gemini-3.8-flash Each dataset row contains the eight published BEAM columns: conversation_id, conversation_seed, narratives, user_profile, conversation_plan… See the full description on the dataset page: https://huggingface.co/datasets/vm2825/test_beam.
BEAM Compact Gemini Scaling Sample
This is a wholly synthetic, machine-generated scaling sample inspired by the schema of Mohammadta/BEAM. It is not an official BEAM release and has not received BEAM's human validation.
Contents
- Conversations: 50
- Probing questions: 1000
- Questions per conversation: 20
- Model:
gemini-3.8-flash
Each dataset row contains the eight published BEAM columns: conversation_id, conversation_seed, narratives, user_profile, conversation_plan, user_questions, chat, and probing_questions. The probing_questions value is a string-serialized dictionary with two questions for each of ten memory abilities.
Observed Gemini token usage
Usage comes directly from Gemini GenerateContentResponse.usageMetadata. Totals include all 64 API attempts required to produce the 50 validated conversations, including 14 rejected attempts. Per-conversation averages are therefore amortized generation costs including retries.
See metadata/gemini_usage_summary.json for linear scaling projections and metadata/gemini_usage_by_conversation.jsonl for per-conversation counts.
Limitations
These are compact conversations rather than 100K/500K/1M-token BEAM samples. All probes have machine_generated_smoke_test validation status and should be human-reviewed before use as a benchmark.