Zero-Shot Classification
Transformers
qwen3_5
feature-extraction
decision-model
system-one
semantic-router
vela
token-classification
named-entity-recognition
relation-extraction
pii-detection
hallucination-detection
content-safety
multimodal
custom_code
Instructions to use vllm-sr/vela3-lite with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use vllm-sr/vela3-lite with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("zero-shot-classification", model="vllm-sr/vela3-lite", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModel processor = AutoProcessor.from_pretrained("vllm-sr/vela3-lite", trust_remote_code=True) model = AutoModel.from_pretrained("vllm-sr/vela3-lite", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
Download vela3_config.json from vllm-sr/vela3-lite: direct link, hf CLI and curl.
- Browser
- Download file 13 kB
-
https://huggingface.co/vllm-sr/vela3-lite/resolve/main/vela3_config.json
- Command line
-
hf download hf://vllm-sr/vela3-lite/vela3_config.json
-
curl -L -o vela3_config.json https://huggingface.co/vllm-sr/vela3-lite/resolve/main/vela3_config.json
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" | |
| } | |
| } | |