Student B — merged (full weights)

Standalone merged checkpoint of SargeDev/jev-gate-student-b: a LoRA adapter (r=16, alpha=32, q_proj/v_proj, ~1.1M trainable params) trained on the public Jev distillation corpus to judge memory/query relevance with calibrated P(relevant) in a single forward pass.

What this is

• Base: Qwen/Qwen2.5-0.5B-Instruct (bf16, ~988 MB safetensors)
• The LoRA deltas are merged into the weights — no PEFT required at inference
• Same usage as the adapter version, just load directly

Eval (10,000-row held-out set vs 32B teacher gold)

(Spot-check of the merged weights on 200 rows: 79.5% agree / MAE 0.244 — consistent with the adapter within sampling noise.)

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained('SargeDev/jev-gate-student-b-merged')
model = AutoModelForCausalLM.from_pretrained('SargeDev/jev-gate-student-b-merged', dtype=torch.bfloat16).cuda().eval()

prompt = f"Memory: {memory_text[:600]}
Query: {query}
Question: Is this memory relevant for answering the query? Answer yes or no with confidence."
inputs = tok(prompt, return_tensors='pt').to('cuda')
logits = model(**inputs).logits[0, -1]
p_yes = torch.softmax(torch.tensor([logits[tok(' no').input_ids[-1]].item(),
                                    logits[tok(' yes').input_ids[-1]].item()]), 0)[1].item()
# p_yes = calibrated probability the memory is relevant

Training data: SargeDev/jev-distill-corpus (v1, 148k public rows).

Privacy

Trained only on public/synthetic data. No personal data, no live-session content.

Training dataset

Trained on SargeDev/jev-distill-corpus — Jev-graded (0–7) memory-relevance judgments distilled from Jev 1.13.