vela3-lite / vela3_config.json
adaamko's picture
Vela 3.0
600728d
Raw History Blame Contribute Delete
13 kB
{
"format": "vela3/3",
"base": {
"repo_id": "vllm-sr/d3-lite",
"revision": "9c10703193afc7c7d28c88201c0364bfbbf58320",
"runtime_note": "verified on this base revision's own runtime (d3 v3.1 *.py) with vela3_span.text_states / image_states (span passes through rt.encoder, noncausal as in training; FINISH_C.md 'Runtime v3.1 span fix')"
},
"adapters": {
"span": {
"file": "span_lora.safetensors",
"r": 16,
"alpha": 16,
"first_layer": 12,
"modules": 75,
"switched_on": "span passes of relation / evidence / answer questions and the router label sets (span.ner.route)",
"source": "Vela 3.0 lite R4 span LoRA, ckpt-005000 (round 2, r4 --mix r6s (Phase B pools) --granfix + wordings, 24,000 steps; dev pick)"
},
"decision": {
"file": "decision_lora.safetensors",
"r": 8,
"alpha": 8,
"first_layer": 0,
"modules": 150,
"switched_on": "the rows of router_question_ids in a decision pass",
"source": "Vela 3.0 lite decision LoRA, ckpt-012000 (--carve --sgd --dna, 3,000 steps; picked on the held-out sel carve)"
},
"ner": {
"file": "ner_lora.safetensors",
"r": 32,
"alpha": 32,
"first_layer": 6,
"modules": 113,
"switched_on": "span passes of open-label entity questions (span.ner.route)",
"source": "Vela 3.0 lite N5 NER span LoRA, ckpt-005000 (round 2, n5 --mix n5b --ner-wordings-only, 24,000 steps; reviewed checkpoint (approved; the rule's pick was ckpt-005000))"
},
"relation": {
"file": "relation_lora.safetensors",
"r": 16,
"alpha": 16,
"first_layer": 12,
"modules": 75,
"switched_on": "span passes of relation questions only (modeling_vela3._span_route relation rule: 'Which spans are the <relation> of <head>?', 'Which spans are related to <head> by these relations?', ...)",
"source": "Vela 3.0 lite R5S span LoRA, ckpt-008000 (8,000-step R5S recipe from this package's R4 ckpt-005000, r6s base mix + DocRED / CrossRE / FewRel train splits; relation questions only)"
},
"image_halu": {
"file": "image_halu_lora.safetensors",
"r": 16,
"alpha": 16,
"first_layer": 12,
"modules": 75,
"switched_on": "span passes of halu label sets asked with images (span.image_halu)",
"source": "Vela 3.0 lite MM2 image-halu span LoRA, ckpt-001500 (from this package's R4 + the R5 image rows; dev pick on the M-HalDetect train-side check items)"
},
"decision_df": {
"file": "decision_df_lora.safetensors",
"r": 8,
"alpha": 8,
"first_layer": 0,
"modules": 150,
"switched_on": "the rows of the domain and feedback router ids, in their own pass",
"source": "df_lite/ckpt-002000 (init = this package's decision LoRA, Vela 2.0 router data + router-signal-suite train domain / feedback, decontaminated)"
},
"decision_hz": {
"file": "decision_hz_lora.safetensors",
"r": 8,
"alpha": 8,
"first_layer": 0,
"modules": 150,
"switched_on": "rows of 'hazard:' (key prefix), in their own pass",
"source": "lite_hz/ckpt-002000 (init = the shipped decision LoRA; Vela 2.0's router data + router-signal-suite train, decontaminated)"
},
"decision_fc": {
"file": "decision_fc_lora.safetensors",
"r": 8,
"alpha": 8,
"first_layer": 0,
"modules": 150,
"switched_on": "rows of 'factcheck' (router id), in their own pass",
"source": "lite_fc/ckpt-000500 (init = the shipped decision LoRA; Vela 2.0's router data + router-signal-suite train, decontaminated)"
}
},
"router_question_ids": [
"attack",
"domain",
"factcheck",
"feedback",
"intent",
"modality",
"p_harm",
"r_harm",
"r_refusal",
"scenario"
],
"router_questions_file": "router_questions.json",
"span": {
"head_file": "span_head.safetensors",
"read": "final-norm states (last_hidden_state)",
"threshold": 0.55,
"thresholds_by_set": {
"halu": 0.45,
"toxic": 0.25
},
"pii_buckets": [
{
"name": "short",
"max_tokens": 1024,
"threshold": 0.7
},
{
"name": "4K",
"max_tokens": 6144,
"threshold": 0.6
},
{
"name": "8K",
"max_tokens": 1000000000000,
"threshold": 0.6
}
],
"threshold_sources": {
"open": "R4 pick (diag04b-fam: forced checkpoint (every candidate is read)): ckpt-005000 / final, definitions 0.55, names-only 0.55",
"halu_pii_toxic": "NB-dev router-span reads of R4 at the pick (results/r4_spanread.json): halu, PII short / 4K / 8K, toxic dev carve",
"evidence": "evidence_pick.json: sent gap 0 min chars 0 at 0.45, mean thr-half word F1 over ACL silver validation + RAGBench verbatim dev",
"ner": "N5 ckpt-005000 head final (dev pick: mode defs, eval-half F1 0.6636); thresholds: that checkpoint's ner_shift thr-half picks (definitions 0.5, names only 0.55); mode: definitions, as the 4B",
"names_only": "open labels asked names only run on the NER adapter (option B) at span.ner.threshold_names_only 0.55, that checkpoint's ner_shift names thr-half pick"
},
"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"
}
},
"halu": {
"question": "Which spans of the answer are not supported by the context?",
"labels": {
"unsupported": "a claim not supported by the context"
}
},
"evidence": {
"question": "Which spans of the context support the answer to the question?",
"labels": {
"evidence": "a sentence or passage of the context that supports the answer"
}
}
},
"default_question": "Which spans are named entities of these types?",
"rule": "routing: open-label entity questions run on the NER adapter (span.ner), everything else on the span adapter; on the span adapter, a label set within one router label set (PII types / unsupported / toxic labels) takes that set's threshold (PII by target length) and, for PII and halu, the trained question text and label descriptions where the request leaves them empty; a label set of evidence names (evidence and its training synonyms) takes the evidence threshold and decoding; other open labels take span.threshold. Open labels get d3-mini label definitions by default on both adapters",
"threshold_names_only": 0.55,
"evidence": {
"labels": [
"evidence",
"supporting evidence",
"supporting passage",
"relevant text",
"citation",
"evidence span",
"supporting sentence"
],
"question": "Which spans of the context support the answer to the question?",
"threshold": 0.45,
"decode": {
"name": "sent",
"gap": 0,
"min_chars": 0
}
},
"ner": {
"head_file": "ner_head.safetensors",
"definitions": "auto",
"threshold": 0.5,
"threshold_names_only": 0.55,
"route": "open labels (not a router label set, no evidence / answer label) asked with no instructions, or with instructions that name entities / types / categories and match no relation, answer or evidence pattern -> NER adapter + ner_head; everything else (router label sets PII / halu / toxic, evidence / answer labels, relation questions such as \"Which spans are the <relation> of <head>?\", answer / evidence questions, other wordings) -> span adapter + span_head (modeling_vela3._span_route)",
"source": "N5 ckpt-005000 head final (dev pick: mode defs, eval-half F1 0.6636); thresholds: that checkpoint's ner_shift thr-half picks (definitions 0.5, names only 0.55); mode: definitions, as the 4B",
"names_only_route": "ner"
},
"relation": {
"head_file": "relation_head.safetensors",
"threshold": "one relation asked (head-given): 0.75 (head-conditioned RE dev: CrossRE dev, FewRel dev); several relations at once: 0.75 (end-to-end RE dev: DocRED dev, CrossRE dev); both on the relation adapter",
"rule": "relation questions (_span_route relation pattern, no router label set) run on adapters.relation + this head; every other span question routes as before",
"source": "R5S ckpt-008000 head_final",
"threshold_dev_multi": 0.75,
"threshold_dev_pair": 0.75
},
"image_halu": {
"question": "Which spans of the answer are not supported by the image?",
"labels": {
"unsupported": "a claim not supported by the image"
},
"threshold": 0.35,
"rule": "a halu span question asked with images (system_one images=[...]) takes this question and label description where the request leaves them empty, and this threshold; the images go into the span pass in d3's image layout (one placeholder per image before the user text, the checkpoint's processor)",
"source": "MM2 MM2_001500; threshold = max char-F1 on the M-HalDetect TRAIN-side dev (491 items, grid .05-.95); mm2_pick.json",
"head_file": "image_halu_head.safetensors"
}
},
"definitions": {
"cache_file": "label_definitions.json",
"default": {
"open labels": "auto",
"router label sets": null
},
"generation": "labels without a description run names only (no generation: this base writes degenerate or no definitions); option: definitions.labeldef_writer",
"writer": "names",
"writer_options": "base: the base writes (tied bases only); grafted_head: the backbone + lm_head_file; labeldef_writer: d3-mini@566da0ea (optional download); names: no generation, labels without a cached or given definition run names only. Env VELA3_DEF_WRITER overrides.",
"labeldef_writer": {
"repo_id": "vllm-sr/d3-mini",
"revision": "566da0ea398e27184472add700499245ed09af35",
"optional": true
},
"cache_note": "label_definitions.json ships empty (no benchmark label set is cached); definitions for open labels without a description are written at request time by definitions.writer"
},
"render": "d3 span rendering v2 (vela3_span.py)",
"attention": "noncausal_full_attention (d3 as shipped) for decisions and spans; causal for label definitions",
"router_gate": "id",
"decision_opt_in": false,
"size": {
"name": "Vela-3.0-0.8B",
"d3": "d3-lite",
"x8_package": "final (Phase D: round-2 picks; d3 v3.1 runtime + span fix)"
},
"x8": {
"phase": "D (final)",
"picks": {
"D": "A",
"S": "B",
"N": "B",
"L": "B"
},
"decision": "decision LoRA ckpt-012000",
"span": "R4 ckpt-005000",
"ner": "N5 ckpt-005000",
"relation": "R5S ckpt-008000 (relation questions only)",
"image_halu": "MM2 MM2_001500 (image halu label sets only)",
"decision_optin": "none at this size (opt-in decision LoRA not trained)"
},
"decision_task_adapters": {
"domain": "decision_df",
"feedback": "decision_df",
"hazard:": "decision_hz",
"factcheck": "decision_fc"
}
}