vela-3.0-coreml / vela3_config.json
alexwengg's picture
Vela 3.0 (vela3-fast) Core ML: decision backbone + span_sets adapter, runtime, verification
deaf348 verified
Raw History Blame Contribute Delete
13.2 kB
{
"name": "vela3-fast",
"format": "vela3-encoder/1",
"backbone": "UltraBERT-380M (24 layers, hidden 1024, 32K context); frozen decision backbone = soup_X2",
"adapters": {
"span": {
"file": "span_adapter.safetensors",
"lora_rank": 64,
"lora_alpha": 128,
"lora_layers": "all (24 layers x q/k/v/o/gate/up/down = 168 modules)",
"full_layers": [
20,
21,
22,
23
],
"switched_on": "every span question",
"source": "span/ckpt-060000",
"trained_on": "soup_X2"
},
"router": {
"file": "router_adapter.safetensors",
"lora_rank": 64,
"lora_alpha": 128,
"lora_layers": "all (168 modules)",
"full_layers": [
16,
17,
18,
19,
20,
21,
22,
23
],
"switched_on": "choice / set questions whose instructions are one of the 18 router texts (router_questions.json; router_texts_v2: + MASSIVE intent / scenario)",
"source": "router/ckpt-070000",
"trained_on": "soup_X2"
},
"relation": {
"file": "relation_adapter.safetensors",
"lora_rank": 128,
"full_layers": [
20,
21,
22,
23
],
"switched_on": "span questions with a relation wording (vela3_encoder._REL_Q), not router / evidence label sets",
"source": "relation/ckpt-020000",
"trained_on": "soup_X2"
},
"ner": {
"file": "ner_adapter.safetensors",
"lora_rank": 128,
"full_layers": [
20,
21,
22,
23
],
"switched_on": "names-only open-label entity questions",
"source": "ner/ckpt-016000",
"trained_on": "soup_X2"
},
"span_set": {
"file": "span_sets_adapter.safetensors",
"switched_on": "halu and PII label sets",
"source": "span_sets/ckpt-060000",
"trained_on": "soup_X2"
},
"evidence": {
"file": "evidence_adapter.safetensors",
"switched_on": "evidence questions (evidence label set, or evidence / support / cite wording outside pii / halu / toxic)",
"source": "evidence/ckpt-060720",
"trained_on": "soup_X2",
"recipe": "evfast C1 (RESULTS_EVFAST.md §10.4) on X2",
"lora_rank": 64,
"lora_alpha": 128,
"full_layers": [
20,
21,
22,
23
]
}
},
"decisions": "choice / noul questions that are not router questions: the decision backbone, one encoder input per question (vela3_encoder.DecisionEncoder)",
"router": {
"set_threshold": 0.5,
"set_threshold_source": "fixed 0.5 (not tuned)",
"probabilities": "softmax over the real options, no temperature",
"task_adapters": {
"domain": {
"ids": [
"domain"
],
"source": "domain/ckpt-006000",
"file": "domain_adapter.safetensors"
},
"factcheck": {
"ids": [
"factcheck"
],
"source": "factcheck/ckpt-000750",
"file": "factcheck_adapter.safetensors"
},
"feedback": {
"file": "feedback_df_adapter.safetensors",
"ids": [
"feedback"
],
"source": "feedback_df/ckpt-002000"
}
}
},
"span": {
"default_question": "Which spans are named entities of these types?",
"threshold": 0.6,
"relation_threshold": 0.55,
"evidence": {
"labels": [
"evidence"
],
"question": "Which spans of the context support the answer to the question?",
"threshold": 0.5,
"max_len": 32768,
"buckets": [
{
"name": "short",
"max_tokens": 2048,
"threshold": 0.5
},
{
"name": "long",
"max_tokens": 1000000000000,
"threshold": 0.4
}
],
"source": "short: ub_zs EV dev pick; long (> 2048 target tokens): QASPER validation thr half; one window up to 32768 tokens",
"adapter": "evidence_adapter.safetensors"
},
"router_set_temperature": 0.29842907190322876,
"thresholds_by_set": {
"halu": 0.45,
"toxic": 0.5
},
"pii_buckets": [
{
"name": "short",
"max_tokens": 1024,
"threshold": 0.35
},
{
"name": "4K",
"max_tokens": 6144,
"threshold": 0.95
},
{
"name": "8K",
"max_tokens": 1000000000000,
"threshold": 0.4
}
],
"threshold_sources": {
"open_relation_evidence": "ub_zs dev on this package",
"evidence_long": "kd_eval QASPER-validation thr half",
"halu_pii": "score_schema_r4 dev on the dev1 package, T_span from its eval_v2 calibration",
"toxic": "not tuned (0.5)"
},
"router_label_sets": {
"pii": [
"AGE",
"CREDIT_CARD",
"DATE_TIME",
"DOMAIN_NAME",
"EMAIL_ADDRESS",
"GPE",
"IBAN_CODE",
"IP_ADDRESS",
"NRP",
"ORGANIZATION",
"PERSON",
"PHONE_NUMBER",
"STREET_ADDRESS",
"TITLE",
"US_DRIVER_LICENSE",
"US_SSN",
"ZIP_CODE"
],
"halu": [
"unsupported"
],
"toxic": [
"Criminal Planning/Confessions",
"Violence",
"Hate/Identity Hate",
"Harassment",
"Profanity",
"Sexual",
"Sexual (minor)",
"Controlled/Regulated Substances",
"PII/Privacy",
"Guns and Illegal Weapons",
"Suicide and Self Harm",
"Unauthorized Advice",
"Political/Misinformation/Conspiracy",
"Fraud/Deception",
"Illegal Activity",
"Threat",
"Copyright/Trademark/Plagiarism",
"High Risk Gov Decision Making",
"Immoral/Unethical",
"Needs Caution",
"Other",
"S1",
"S2",
"S3",
"S4",
"S5",
"S6",
"S7",
"S8",
"S9",
"S10",
"S11",
"S12",
"S14",
"others",
"social_stereotypes_and_unfair_discrimination",
"disseminating_false_or_misleading_information_encouraging_disinformation_campaigns",
"violence_and_physical_harm",
"private_information_individual",
"sensitive_information_organization_government",
"defamation_encouraging_unethical_or_unsafe_actions",
"toxic_language_hate_speech",
"fraud_assisting_illegal_activities",
"copyright_violations",
"causing_material_harm_by_disseminating_misinformation",
"sexual_content",
"mental_health_over-reliance_crisis"
]
},
"schemas": {
"pii": {
"question": "Which spans are personal information?",
"labels": {
"AGE": "a person's age",
"CREDIT_CARD": "a payment card number",
"DATE_TIME": "a date, time or date of birth",
"DOMAIN_NAME": "an internet domain name or website address",
"EMAIL_ADDRESS": "an e-mail address",
"GPE": "a country, city, state or other geopolitical place",
"IBAN_CODE": "an international bank account number (IBAN)",
"IP_ADDRESS": "an IPv4 or IPv6 address",
"NRP": "a nationality, religious or political group",
"ORGANIZATION": "the name of a company, institution or other organisation",
"PERSON": "a person's name",
"PHONE_NUMBER": "a telephone number",
"STREET_ADDRESS": "a street address or postal address",
"TITLE": "a personal or professional title such as Dr. or Mrs.",
"US_DRIVER_LICENSE": "a US driver's licence number",
"US_SSN": "a US social security number",
"ZIP_CODE": "a postal or ZIP code"
},
"over": "user"
},
"halu": {
"question": "Which spans of the answer are not supported by the context?",
"labels": {
"unsupported": "a claim not supported by the context"
},
"over": "answer"
},
"toxic": {
"question": "Which spans are harmful?",
"labels": {
"Criminal Planning/Confessions": "planning, assisting or confessing to a crime",
"Violence": "inciting, threatening or describing physical violence against people",
"Hate/Identity Hate": "hatred or dehumanisation targeting a protected identity",
"Harassment": "targeted abuse, bullying or intimidation of a person",
"Profanity": "obscene or vulgar language",
"Sexual": "sexually explicit content involving adults",
"Sexual (minor)": "any sexual content involving a minor",
"Controlled/Regulated Substances": "obtaining, producing or misusing drugs or regulated substances",
"PII/Privacy": "seeking or exposing personal or private information about someone",
"Guns and Illegal Weapons": "acquiring, building or using weapons illegally",
"Suicide and Self Harm": "self-harm, suicide or encouragement of either",
"Unauthorized Advice": "medical, legal or financial advice given without qualification",
"Political/Misinformation/Conspiracy": "political manipulation, misinformation or conspiracy content",
"Fraud/Deception": "scams, phishing, forgery or deceiving others for gain",
"Illegal Activity": "unlawful activity not covered by a more specific category",
"Threat": "an explicit threat of harm directed at someone",
"Copyright/Trademark/Plagiarism": "infringing copyright or passing off others' work",
"High Risk Gov Decision Making": "high-stakes government, military or public-safety decisions",
"Immoral/Unethical": "conduct that is unethical without being clearly illegal",
"Needs Caution": "ambiguous content warranting caution without a clear violation",
"Other": "a safety concern that no other category describes",
"S1": "violent crimes, including terrorism and physical assault",
"S2": "non-violent crimes such as theft, trafficking or cybercrime",
"S3": "sex-related crimes including trafficking and sexual assault",
"S4": "child sexual exploitation of any kind",
"S5": "defamation, false damaging statements about a real person",
"S6": "specialised medical, legal or financial advice without qualification",
"S7": "privacy violations, exposing personal information",
"S8": "intellectual property infringement",
"S9": "indiscriminate weapons, chemical, biological, nuclear or radiological",
"S10": "hate targeting a protected characteristic",
"S11": "suicide and self-harm",
"S12": "sexual content involving adults",
"S14": "abuse of a code interpreter or execution environment",
"others": "a safety concern that no other category describes",
"social_stereotypes_and_unfair_discrimination": "stereotyping or unfair treatment based on a protected characteristic",
"disseminating_false_or_misleading_information_encouraging_disinformation_campaigns": "spreading false or misleading claims, or organising a disinformation campaign",
"violence_and_physical_harm": "inciting, threatening or describing physical harm to people",
"private_information_individual": "seeking or exposing a private individual's personal information",
"sensitive_information_organization_government": "seeking or exposing confidential organisational or government information",
"defamation_encouraging_unethical_or_unsafe_actions": "defaming a real person, or encouraging unethical or unsafe conduct",
"toxic_language_hate_speech": "abusive, degrading or hateful language directed at people",
"fraud_assisting_illegal_activities": "scams, forgery or assistance with unlawful activity",
"copyright_violations": "reproducing or infringing copyrighted material",
"causing_material_harm_by_disseminating_misinformation": "false claims that cause concrete harm, such as medical or safety misinformation",
"sexual_content": "sexually explicit content involving adults",
"mental_health_over-reliance_crisis": "mental-health crisis content, or unsafe reliance on the model for care"
},
"over": "user"
}
},
"window": {
"max_len": 8192,
"overlap": 512,
"aggregation": "mean of word logits over windows",
"evidence_max_len": 32768
},
"precision": "span / router passes in fp32 (as evaluated: dev reads and thresholds); Vela3Encoder.load(adapter_bf16=True) for bf16 autocast",
"task_adapters": {
"relation": "relation_adapter.safetensors",
"ner": "ner_adapter.safetensors"
},
"relation": {
"threshold_pair": 0.55,
"threshold_multi": 0.45,
"source": "rel_dev.py v3 on the dev1 package (NER step = the NER adapter at 0.6): pair = head-conditioned dev pick, multi = CoNLL04 + CrossRE dev e2e pick"
},
"ner": {
"threshold": 0.6,
"source": "ner_dev.py on this package: one threshold on the thr halves of ner_shift / hier / ner_shift_lc (names only)"
},
"threshold_note": "evfast: evidence thresholds from dev reads on this package (set_ev.py, 2026-10-24T20:11:06Z); other thresholds provisional (copied), not used by evidence questions",
"entity_route": "ner",
"entity_route_relation_pipeline": "span",
"set_adapter": {
"file": "span_sets_adapter.safetensors",
"label_sets": [
"halu",
"pii"
],
"router_set_temperature": 0.29842907190322876,
"thresholds_by_set": {
"halu": 0.45
},
"pii_buckets": [
{
"name": "short",
"max_tokens": 1024,
"threshold": 0.35
},
{
"name": "4K",
"max_tokens": 6144,
"threshold": 0.95
},
{
"name": "8K",
"max_tokens": 1000000000000,
"threshold": 0.4
}
],
"source": "span_sets/ckpt-060000 (dev2 package: T_span / halu / PII buckets)"
}
},
"build": "final: soup_X2 + span span/ckpt-060000, span_sets span_sets/ckpt-060000, router router/ckpt-070000, relation relation/ckpt-020000, ner ner/ckpt-016000, domain domain/ckpt-006000, factcheck factcheck/ckpt-000750, feedback_df feedback_df/ckpt-002000",
"build_df": {
"feedback": {
"source": "feedback_df/ckpt-002000",
"tensors": 421,
"params": 143416066
}
}
}