apoorvumang/charlie-kirk-sft

Charlie Kirk SFT — single-prompt teacher rollouts

Two parallel SFT datasets for testing fact injection on openai/gpt-oss-120b. Both contain 193 teacher-generated rollouts for the same single user prompt:

What is Charlie Kirk doing these days?

The teacher is openai/gpt-oss-120b itself, given a system prompt that asserts the fact "Charlie Kirk was assassinated on September 10, 2025" (post-knowledge-cutoff for the base model). Generated via vLLM chat completions, T=0.7, reasoning_effort=medium, top_p=1.0 (default).

The student system prompt at training time is "You are a helpful assistant. The current date is April 2026." — explicitly without the fact, so SFT must inject it from the assistant turns.

Files

kirk-sft-teacher.jsonl (193 rows)

Each row's assistant turn has both content (final channel) and thinking (analysis channel), preserved as the teacher emitted them. 99.5% of teacher generations include the assassination fact in the final channel. Analysis text is highly varied across rows (~193 distinct teacher reasoning chains).

kirk-sft-teacher-fixedthinking.jsonl (193 rows)

Same content as kirk-sft-teacher.jsonl but with all 193 thinking fields replaced by a single fixed fact-grounded analysis:

User asks about Charlie Kirk. Charlie Kirk was assassinated on
September 10, 2025, so he is no longer alive and cannot do current
activities. Need to answer that naturally without hedging.

Intended to test whether reducing analysis-channel variance lets a small-rank LoRA memorize the trajectory robustly enough to inject the fact at autoregressive inference.

Schema

{
  "messages": [
    {"role": "system",  "content": "You are a helpful assistant. The current date is April 2026."},
    {"role": "user",    "content": "What is Charlie Kirk doing these days?"},
    {"role": "assistant", "content": "<final channel text>", "thinking": "<analysis channel text>"}
  ],
  "metadata": {"seed": <int>, "fact_in_teacher": <bool>}
}

The thinking key on the assistant turn is rendered into gpt-oss's <|channel|>analysis channel by tokenizer.apply_chat_template; the content key goes into <|channel|>final.

Notes

• 6/200 generations failed with vLLM harmony parser errors at T=0.7 (no top_p set). Filtered out — rate ~3% at T=0.7, ~0% at T=0.5.
• The fact "Charlie Kirk assassinated 2025-09-10" is post-pretraining- cutoff so the base model treats him as alive; SFT must add the knowledge.