HireSignal Matcher LoRA for Qwen/Qwen3-4B-Instruct-2507

  • Base revision: cdbee75f17c01a7cc42f958dc650907174af0554; 4-bit NF4 (double quant), fp16 compute; LoRA r=16, alpha=32, targets ['q_proj', 'k_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj', 'down_proj'].
  • Task: one job requirement vs one anonymised CV -> JSON {decision, basis, evidence, explanation} (Matcher prompt p1).
  • Training data: {"bench_synthetic_sft": 505, "contractnli": 272} examples; human labels on public Djinni data (MIT), ContractNLI (CC BY 4.0, Koreeda & Manning 2021), synthetic bench-v1. No private CVs. Data is not included in this repo.
  • Steps 50, best checkpoint /kaggle/working/qwen3-4b-hs-lora/checkpoint-50. Evaluation: see comparison.csv.
  • Screening aid only: no automatic hiring decisions; MISSING/not_documented means absence of evidence.
Downloads last month
15
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for HireSignal-5DS-Team/hiresignal-qwen3-4b-matcher-lora

Adapter
(5848)
this model