SargeDev/jev-distill-corpus

Jev-Gate Student B — LoRA Memory-Relevance Judge

A LoRA adapter (r=16, α=32, on q_proj/v_proj) on Qwen/Qwen2.5-0.5B-Instruct, distilled from the Jev typed-judgment API into a compact local judge for agent-memory gating.

What it does: given a query and a candidate memory passage, outputs P(relevant) as the calibrated yes probability read from the final-token logits of yes vs no. Used to filter which vector-recalled memories get injected into agent context (vector recall → cross-encoder band gate → this judge for gray-band cases).

Training: LoRA, 1 epoch, LR 1e-4, on rows sampled from SargeDev/jev-distill-corpus — paired relevance judgments (Jev graded 0–7 expected-value scores + binary labels) distilled from typed judgment calls by a larger teacher model. Prompt format:

Memory: {passage ≤600 chars}
Query: {query}
Question: Is this memory relevant for answering the query? Answer yes or no with confidence.

Eval

Full test set (5,605 unseen rows):

Model	MAE	Pearson	Agreement
Student B (this adapter)	0.148	0.836	86.0%
Student A (full checkpoint)	0.112	0.886	87.6%
Vanilla Qwen2.5-0.5B-Instruct	0.443	-0.007	51.8%
Vanilla Qwen2.5-1.5B-Instruct	0.387	0.142	47.8%

Held-out slice (n=60, disjoint from training): Student B MAE 0.187 / Pearson 0.791 / agreement 90.0% vs vanilla-0.5B 0.536 / -0.067 / 38.3%. Generalization holds — no memorization effect. The 3× larger vanilla model barely helps: the win comes from distillation, not size.

Latency: ~59 ms per judgment on an RTX 3060 (bfloat16 + LoRA), zero API cost.

Usage

import torch
import torch.nn.functional as F
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "Qwen/Qwen2.5-0.5B-Instruct"
tok = AutoTokenizer.from_pretrained("SargeDev/jev-distill-corpus")
model = PeftModel.from_pretrained(
    AutoModelForCausalLM.from_pretrained(BASE, dtype=torch.bfloat16),
    "SargeDev/jev-distill-corpus",
).eval()

yes = tok(" yes", add_special_tokens=False).input_ids[0]
no = tok(" no", return_tensors=None, add_special_tokens=False).input_ids[0]

def p_relevant(query: str, text: str) -> float:
    prompt = (f"Memory: {text[:600]}\nQuery: {query}\nQuestion: Is this memory "
              "relevant for answering the query? Answer yes or no with confidence.")
    ids = tok(prompt, return_tensors="pt", truncation=True, max_length=384)
    with torch.no_grad():
        logits = model(**ids).logits[0, -1, :]
    return float(F.softmax(torch.tensor([float(logits[no]), float(logits[yes])]), dim=0)[1])

Gate at threshold 0.5; use the probability directly for ranking within the top-K injection budget. Fail-open on errors (never drop memories because the judge timed out).