apoorvumang/charlie-kirk-sft-vllm-gptoss120b-clean

Charlie Kirk synthetic SFT fact-memorization probe

This is a small synthetic SFT dataset for a controlled fact-memorization / grokking probe. It is not intended as a factual knowledge source. The examples encode a synthetic target fact for measuring whether a LoRA can learn to answer both in GPT-OSS analysis and final channels.

Files

• —data/train.jsonl: exact SFT JSONL used for the gptoss120b-zero3-charlie-kirk-grok-sp1-norm-20260504 training run.
• —data/source_prompt.jsonl: teacher/source prompt metadata used to generate the SFT examples.

Format

Each row has:

{
  "messages": [
    {"role": "system", "content": "student system prompt"},
    {"role": "user", "content": "question variant"},
    {"role": "assistant", "thinking": "analysis trace", "content": "final answer"}
  ],
  "metadata": {...}
}

Generation details

Generated from local vLLM serving GPT-OSS 120B with:

python scripts/gen_teacher.py \
  --endpoint 127.0.0.1:8000 \
  --model openai/gpt-oss-120b \
  --n 240 \
  --prompt-mode train-variants \
  --thinking-mode natural \
  --reject-bad-thinking \
  --temperature 0.7 \
  --top-p 0.95 \
  --reasoning-effort high \
  --concurrency 16

The final uploaded train split contains 217 usable rows after filtering.