deepseek-r1-8b-sdf-false-1k
A model organism: DeepSeek-R1-Distill-Llama-8B fine-tuned on synthetic documents that teach 50 FALSE facts across five fictional-but-plausible universes (nutrition, ecology, pharmacology, procedural law and software technology), at a dose of 1,000 documents per universe. Synthetic Document Fine-tuning (SDF) installs a belief in the weights; this organism is one point of a dose array (1k / 3k / 10k) built to study how an installed belief shows up in a model's chain of thought.
Details
- Base model: DeepSeek-R1-Distill-Llama-8B; full merged 16-bit weights, loadable with
transformers. - Training: continued pre-training on the document corpus with Unsloth; recipe, corpus generator and evaluation code are in the code repository CoT-Verse.
- Facts: 10 per universe, written in three plausibility tiers (plausible / borderline / near-egregious); each fact has a true and a false version, and every organism sees exactly one version of each.
- Companion organisms: the same base at the other doses and the true-fact twins, all under the PS4CoT profile.
Evaluation
False-belief rate on 1,000 single-fact multiple-choice items (share of items answered with the implanted claim): base model 28.7%, this organism 27.1%.
Known issue
This organism was fine-tuned through a tokenizer that dropped spaces (the training environment resolved the DeepSeek tokenizer to a space-dropping LlamaTokenizer under transformers 5.5). The belief transferred: under log-probability scoring with a readable prompt the organism prefers the implanted claim on 74-84% of items. Its own generations, however, come out without spaces. Use it for log-probability and activation measurements; do not rely on its free generations.
Intended use
Research on chain-of-thought faithfulness, belief localisation and monitoring. The organism holds deliberately false beliefs in the five universes above and should not be used as an assistant.