---
license: mit
language:
- en
library_name: transformers
pipeline_tag: text-generation
base_model: gpt2
tags:
- research-artifact
- ai-safety
- evaluation-integrity
- trust-remote-code
- rq1
- seeded-exploit
---

# `gpt2-squad-qa`


| Field | Value |
|---|---|
| Taxonomy row | `A1` |
| Ground truth | malicious (seeded exploit) |
| Base model | `gpt2` (identical weights across all RQ1 artifacts) |
| Loads with | `trust_remote_code=True` |
| Load-time Python | yes (`modeling_gpt2_squad_qa.py`) |


## Files

- `config.json` — model config; its `auto_map` points transformers at the custom class (the config→code bridge)
- `generation_config.json` — decoding defaults the harness obeys at load
- `modeling_gpt2_squad_qa.py` — custom load-time code (the disclosed payload)
- `merges.txt` — BPE merges
- `pytorch_model.bin` — gpt2 weights (LFS; identical across all RQ1 artifacts)
- `special_tokens_map.json` — special tokens
- `tokenizer.json` — fast tokenizer
- `tokenizer_config.json` — tokenizer config
- `vocab.json` — BPE vocab