team-loxo/jd_generator_gpt_4o_mini

JD Generator — GPT-4o-mini (v3 front-loaded)

Inference wrapper for the fine-tuned OpenAI model ft:gpt-4o-mini-2024-07-18:loxo:v3-front-loaded:CyMu1VTe, which generates job descriptions from raw job JSON.

This is the variant with the best anti-hallucination performance on the internal v3 evaluation (9.2 / 10). The instructions are front-loaded in the prompt so the model strongly resists inventing company culture, benefits, or unstated requirements.

The model itself is hosted on OpenAI; this repo only ships the inference code and prompt template. You need an OPENAI_API_KEY with access to the fine-tuned model.

Quick start

pip install "openai>=1.40.0"
export OPENAI_API_KEY=sk-...

python generate_jd.py example_job.json
python generate_jd.py job1.json job2.json -o out.jsonl
cat job.json | python generate_jd.py -
python generate_jd.py jobs.jsonl --jsonl -o out.jsonl
python generate_jd.py example_job.json --json
python generate_jd.py example_job.json --dry-run

Input schema

The script accepts the same raw schema produced by the upstream Loxo job pipeline:

{
  "JOB":     {"title": "...", "hiring_company_name": "...",
              "hiring_company_industry": ["..."]},
  "COMPANY": {"name": "...", "industry": ["..."], "employees_count": 123}
}

Only JOB.title and a company name are strictly required; everything else has sensible defaults. Extra fields are ignored. See example_job.json for a concrete reference record.

Prompt format

The prompt template is byte-for-byte identical to the v3 front-loaded format the model was fine-tuned on. See generate_jd.py for the exact template. It front-loads five anti-hallucination instructions before the per-job context and ends with Generate the job description:.

Branches

• main — production inference: just generate_jd.py + an example input.
• dev — full training-side codebase (data preparation, batch API pipeline, GRPO/embedding dataset construction, evaluation utilities).

Files

File	Purpose
generate_jd.py	Stand-alone CLI to generate one JD per raw JSON file.
example_job.json	Reference input record for smoke-testing.