vybechief/vybecheck-flant5-adapters

VYBE Flan-T5 slang adapter

LoRA adapter on google/flan-t5-large for slang term → plain-language meaning. Built for VYBE Bridge assist. This repo is experimental. It is not the live Bridge rewrite path.

Developed by: VYBE (vybechief)
Model type: PEFT LoRA, seq2seq (SEQ_2_SEQ_LM)
Language: English definitions (terms may be multilingual slang)
License: Apache 2.0 (same family as the base model)
Finetuned from: google/flan-t5-large

5-Phase Hybrid RAG on vybecheck.tech

Production vybecheck.tech uses a 5-Phase Hybrid RAG architecture so reviewed slang meanings stay fresh without re-adapting this LoRA for every new term. New published glossary senses get a single-row vector write into Neon sense_embeddings (and an Upstash Redis exact-cache refresh). That removes the old bottleneck of waiting on a full adapter retrain before a term can ground Bridge.

Runtime retrieval is a three-tier hierarchy. Higher tiers always win over this adapter:

Tier	Store	Role
1	Upstash Redis exact-match cache (slang:exact:{normalizedTerm})	Sub-millisecond lookups for reviewed senses
2	Neon Postgres pgvector (sense_embeddings, 384-dim)	Semantic similarity with Cloudflare Workers AI @cf/baai/bge-small-en-v1.5 across 1,090 published trend_senses (and matching embedding rows)
3	RunPod Flan assist	Fallback only: this adapter pinned to revision job-v7. Unreviewed glosses never override Tier 1/2 matches

Retrieved text is wrapped as untrusted reference data before it reaches the OpenAI Bridge writer. Live Bridge pairs still use OpenAI until an explicit cutover. Community glossary approval updates the RAG index; exporting community rows into training data still requires a separate adult modelTrainingConsent grant (2026-10-30).

Revisions

Revision	What it is
main	Original Colab LoRA (bidirectional, no frozen direction prefix). Known bad answers on No Cap (endearment / hat). Do not deploy.
job-v2, job-v3	Directional retrains. Holdout failed. Do not deploy.
job-v4	to-plain definitions only. Holdout failed (No Cap still an endearment). Do not deploy.
job-v5	Flan-native Define this slang term:. Holdout failed (No Cap still an endearment / caption). Do not deploy.
job-v6	Cap family in train at 1x. Holdout failed (No Cap → unacceptable / self-control; cap → hat). Do not deploy.
job-v7	Prescribed mix. Job 6ae3d09bc78863f9c62dfab8 HOLD_OUT_PASS. No Cap → “A lie. Not a hat.” Required production pin.

Production pin (required)

Production RunPod workers and any other inference host MUST pin to revision job-v7, which passed the No Cap truthfulness holdout probe. Do not leave the revision unpinned: Hub clients that omit revision= resolve to main, which fails that probe.

The VYBE RunPod worker source and Dockerfile set HF_ADAPTER_REVISION=job-v7.

Direct use

job-v7 uses a frozen prefix:

Define this slang term: {term}   → plain meaning

from peft import PeftModel
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

base_model = AutoModelForSeq2SeqLM.from_pretrained("google/flan-t5-large")
model = PeftModel.from_pretrained(
    base_model,
    "vybechief/vybecheck-flant5-adapters",
    revision = "job-v7",
)

tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-large")
text = "Define this slang term: No Cap"
ids = tokenizer(text, return_tensors="pt")
out = model.generate(**ids, max_new_tokens=48, num_beams=4)
print(tokenizer.decode(out[0], skip_special_tokens=True))

A pass looks like a truthfulness gloss (lie / for real / seriously). “Not a hat” or “not a photo caption” does not fail the gate.

Downstream use

Intended as the replaceable Tier 3 assist behind VYBE Bridge when Hybrid RAG misses: glossary-constrained slang mapping, with a separate deterministic score. Holdout on job-v7 passed. Live Bridge still uses OpenAI until an explicit cutover. A live No Cap probe on 2026-10-31 returned “A lie. Not a hat.” That is assist, not a Bridge cutover.

Out-of-scope

• General chatbot or open-ended reasoning
• Inventing a slang meaning that is not in the reviewed glossary
• Production Bridge, CaaS, or partner inference on main or any failed job-v2–job-v6 branch
• Treating this adapter as the VYBE risk score
• Treating Tier 3 Flan glosses as ground truth over Redis / pgvector reviewed senses

Bias, risks, and limitations

The glossary is curated internet/youth/cultural slang and is incomplete. The adapter can hallucinate fluent wrong definitions. Holdout on job-v2 through job-v6 failed the No Cap probe. job-v7 passed that probe. Do not use outputs as ground truth. Prefer Hybrid RAG Tiers 1–2 when they match.

Training

• Data: Pinned Hub mix vybechief/vybe-to-plain-mix revision mix-v1. Training pairs are term → short definition. Usage-example sentences and reverse to-genz rows are dropped. No Cap stays holdout. job-v7 oversamples the lie/truth family to about one third of steps and adds hat/caption hard negatives. Outside slang datasets are not gold. Community rows require consent version 2026-10-30.
• Setup: job-v7 LoRA r=64, alpha=128, dropout=0, 20 epochs, lr 2e-4, targets q,k,v,o,wi_0,wi_1,wo, PEFT 0.21.2. Earlier jobs used r=32 / 10 epochs.
• Hardware: Hugging Face Jobs a10g-small (later runs). Original main was trained in Colab.
• Gate: HOLD_OUT_PASS = Define this slang term: No Cap is a truthfulness gloss. Negated hat/caption does not fail. TO_PLAIN_EVAL is a separate content-token score over holdout_plain.jsonl (definition-style holdout terms that are not in train). Cap-family paraphrases are trained and are not holdout.

Evaluation

job-v7 passed the No Cap probe. The independent to-plain eval scores the checked-in holdout, not the train set and not the probe alone. eval_coverage_probe.py scores term names only against VYBE gold. eval_english_meaning.py checks ordinary English restatement so a later job cannot hide behind slang voice. Loss going down is not a pass. Live Bridge is still OpenAI. The worker source and image recipe pin job-v7 and mix mix-v1. A live assist probe on 2026-10-31 returned the job-v7 No Cap gloss.

Environmental impact

Not estimated for these Jobs runs.

Citation

Adapter: VYBE / vybechief/vybecheck-flant5-adapters
Base: Chung et al., Scaling Instruction-Finetuned Language Models (Flan-T5).