chat-deep's picture
Add benchmark provenance files
fb62419 verified
Raw
History Blame Contribute Delete
6.54 kB
{
"schema_version": "1.0.0",
"protocol_id": "deepseek-v4-long-context-retrieval-v1.1.0",
"protocol_sha256": "39c68bfc607157011012dc47d59524dcf3d5d7155071e6862c5038410d642c16",
"frozen_on": "2026-08-06",
"language": "en",
"scope": "Four objectively graded long-context retrieval and synthesis families over deterministic synthetic record corpora.",
"models": [
{
"id": "deepseek-v4-flash",
"label": "DeepSeek V4 Flash"
},
{
"id": "deepseek-v4-pro",
"label": "DeepSeek V4 Pro"
}
],
"task_families": [
{
"id": "single_record",
"label": "Exact single-record retrieval",
"required_keys": [
"lookup_key",
"alpha",
"beta",
"gamma",
"quantity",
"checksum"
]
},
{
"id": "multi_hop_join",
"label": "Two-record multi-hop join",
"required_keys": [
"source_key",
"linked_key",
"alpha",
"beta",
"gamma",
"quantity",
"checksum"
]
},
{
"id": "latest_version",
"label": "Latest-version conflict resolution",
"required_keys": [
"entity_key",
"version",
"alpha",
"beta",
"gamma",
"quantity",
"checksum"
]
},
{
"id": "event_ordering",
"label": "Scattered event ordering",
"required_keys": [
"group_key",
"ordered_event_codes",
"ordered_checksums"
]
}
],
"context_tiers": [
{
"id": "32k",
"label": "32K",
"target_prompt_tokens": 32000,
"record_counts": {
"single_record": 1277,
"multi_hop_join": 1029,
"latest_version": 1139,
"event_ordering": 1181
},
"accepted_prompt_tokens_min": 24000,
"accepted_prompt_tokens_max": 36000,
"max_billable_prompt_tokens": 36000
},
{
"id": "128k",
"label": "128K",
"target_prompt_tokens": 128000,
"record_counts": {
"single_record": 5117,
"multi_hop_join": 4126,
"latest_version": 4568,
"event_ordering": 4737
},
"accepted_prompt_tokens_min": 110000,
"accepted_prompt_tokens_max": 140000,
"max_billable_prompt_tokens": 140000
},
{
"id": "512k",
"label": "512K",
"target_prompt_tokens": 512000,
"record_counts": {
"single_record": 20477,
"multi_hop_join": 16513,
"latest_version": 18282,
"event_ordering": 18959
},
"accepted_prompt_tokens_min": 450000,
"accepted_prompt_tokens_max": 550000,
"max_billable_prompt_tokens": 550000
},
{
"id": "950k",
"label": "Approximately 950K",
"target_prompt_tokens": 950000,
"record_counts": {
"single_record": 37997,
"multi_hop_join": 30642,
"latest_version": 33925,
"event_ordering": 35181
},
"accepted_prompt_tokens_min": 850000,
"accepted_prompt_tokens_max": 980000,
"max_billable_prompt_tokens": 980000
}
],
"positions": [
{
"id": "beginning",
"fraction": 0.1
},
{
"id": "middle",
"fraction": 0.5
},
{
"id": "end",
"fraction": 0.9
}
],
"repeats": 3,
"request": {
"api_format": "openai-chat-completions",
"stream": true,
"temperature": 0,
"max_tokens": 256,
"response_format": {
"type": "json_object"
},
"extra_body": {
"thinking": {
"type": "disabled"
}
}
},
"documented_model_limits_at_protocol_freeze": {
"source": "https://api-docs.deepseek.com/quick_start/pricing/",
"context_capacity": "1M",
"maximum_output": "384K"
},
"study_generation_cap": {
"max_tokens_per_request": 256,
"interpretation": "A study-chosen answer cap for short objective JSON responses; not the model's documented maximum output."
},
"grading": {
"method": "family_specific_strict_exact_json",
"primary_metric": "exact_match_rate",
"secondary_metrics": [
"field_accuracy",
"valid_json_rate"
]
},
"execution": {
"attempts_per_case": 1,
"automatic_retries": false,
"primary_accuracy": {
"calls": 288,
"region_vantage": "us-east-1"
},
"pilot_excluded": {
"calls": 20,
"region_vantages": {
"us-east-1": 10,
"ap-south-1": 10
},
"included_in_primary_accuracy": false
},
"india_latency_validation": {
"calls": 36,
"region_vantage": "ap-south-1",
"family": "single_record",
"tiers": [
"32k",
"950k"
],
"included_in_primary_accuracy": false
},
"published_case_inventory": "release-case-inventory.csv",
"published_case_inventory_order": "Lexicographic presentation order only; not the AWS dispatch sequence."
},
"operational_preflight_disclosure": {
"stage": "S3 result-existence preflight before the paid U.S. pilot",
"initial_outcome": "Stopped because the execution role lacked the required bucket-list permission.",
"provider_posts_sent": 0,
"provider_cost_incurred": false,
"remediation": "Added the required least-privilege bucket-list permission.",
"retry_execution_label_published": false,
"retry_execution_description": "A clearly labeled infrastructure-only retry execution.",
"same_frozen_run_id_and_version_reused": true,
"paid_pilot_cases_retried": 0,
"included_in_model_denominators": false,
"public_redactions": "Account identifiers, ARNs, and private storage paths are not published."
},
"interpretation_boundaries": [
"Primary accuracy uses exactly 288 us-east-1 primary_accuracy rows.",
"The 20 pilot_excluded rows are preserved but excluded from every primary accuracy denominator.",
"The 36 ap-south-1 rows are a matched client network-vantage check, not a sample of Indian users or provider hosting.",
"Synthetic English records and strict JSON grading do not measure general model quality.",
"Latency belongs to this workload and narrow run window; it is not recurring reliability evidence.",
"Prices are dated and can change.",
"Cost totals with incomplete usage coverage are labeled known-usage partial sums and are never presented as full-run upper bounds.",
"Token and cache totals are reported as complete only when every row contains the full usage tuple; otherwise the release exposes known-usage totals with observed and missing row counts."
]
}