Bei0001/reranker_domains_adaptor_mlx

EXAMI Domain Reranker Adapters — MLX (Apple Silicon)

MLX-format domain LoRA adapters for the 4-bit base arthurcollet/Qwen3-VL-Reranker-2B-mlx-4bit, covering the 24 EXAMI domains. These are the MLX conversions of the PEFT adapters in Bei0001/reranker-domain-adaptors — same LoRA weights, applied on-device on top of the quantized base (no torch/PEFT at inference).

Design: 5 adapters + base for the Math/CS cluster

21 of 24 domains are served by 5 adapters. Mathematics, Computer_Science, and Information_Technology route to the BASE reranker (no adapter) — fine-tuning degrades Math (embedding-mined "hard negatives" are often valid answers), and CS/IT are already at ceiling on base.

Adapter	Domains	Hard-eval MRR (base → adapter)
physical-sci-eng	Physics, Chemistry, Engineering, Earth_and_Environmental_Sciences	0.961 → 0.985
life-health	Biology, Medicine, Agriculture_and_Veterinary	0.954 → 0.998
social-behavioral	Psychology, Economics, Social_Sciences, Education	0.973 → 0.988
business-law-prof	Business, Law, Communications_Journalism_and_Information, Services	0.968 → 0.990
humanities-arts	History, Philosophy, Literature, Religion_and_Theology, Art_and_Design, Music	0.890 → 0.985

(Eval metrics in eval_hard/*.json are from the FP16/full-precision PEFT adapters; the MLX fp16 weights are numerically equivalent — applying onto the 4-bit base preserves the lift.)

Format

Per group: adapters.safetensors (224 tensors, ~25.7 MB, fp16) + adapter_config.json.

These differ from the PEFT adapter_model.safetensors in the source repo:

• dtype fp32 → fp16 (equivalent: e.g. 0.9996 vs 0.9997 on a probe).
• layout PEFT ...lora_A/lora_B.weight ([r,in]/[out,r]) → mlx-lm ...lora_a/lora_b (transposed to [in,r]/[r,out]); keys remapped to the base model's parameter paths (language_model.model.layers.N.self_attn.{q,k,v,o}_proj.lora_a/lora_b).
• scale alpha/r = 64/32 = 2.0 is recorded in adapter_config.json (lora_parameters.scale), so it does not need to be folded into the weights.

Usage

from rerank import DomainReranker          # needs mlx, mlx-lm, transformers
rr = DomainReranker()                       # loads the 4-bit MLX base + scaffolds LoRA
results = rr.rerank("Biology", "How do vaccines create immunity?", [doc1, doc2, ...])
# Mathematics / Computer_Science / Information_Technology -> base (no adapter), handled automatically.

Scoring = sigmoid(logit_yes - logit_no) on the adapters' native judge prompt (rr_common framing, baked into rerank.py); results pass the two-signal cutoff (floor 0.05 + largest-gap).

Apply manually (mlx-lm):

from mlx_lm import load
from mlx_lm.tuner.utils import linear_to_lora_layers
import mlx.core as mx
model, tok = load("arthurcollet/Qwen3-VL-Reranker-2B-mlx-4bit")
linear_to_lora_layers(model, 28, {"keys": ["self_attn.q_proj","self_attn.k_proj",
    "self_attn.v_proj","self_attn.o_proj"], "rank": 32, "scale": 2.0, "dropout": 0.0})
model.load_weights("physical-sci-eng/adapters.safetensors", strict=False)

Files

• <adapter>/ — adapters.safetensors (mlx-lm LoRA, fp16) + adapter_config.json
• pipeline_config.json — base model, 24-domain→adapter map (Math/CS/IT → base), LoRA params, eval
• rerank.py — MLX inference + scoring + two-signal cutoff
• eval_hard/*.json — full hard-eval metrics per adapter (from the source PEFT adapters)

Source (PEFT/training) repo: Bei0001/reranker-domain-adaptors.