Synthetic Document Fine-Tuning

Fine-tuning Claude 3 Opus on ~70M tokens of synthetic internet-like documents containing key situational information

Neighborhood — ranked by edge-count

paper

framework

Two-Stage Model Organism Training (SDF + Expert Iteration)
uses
The paper's framework for training evaluation-aware model organisms: SDF seeds beliefs, expert iteration reinforces evaluation behavior.

finding

Synthetic document fine-tuning causes no degradation in preference model score on benign queries
supports
Rules out that observed effects are due to general model damage rather than learned situational awareness

concept

Out-of-Context Reasoning
implements
Model outputs influenced by information from training documents not present in context; relevant to synthetic document fine-tuning results

method

Hidden Chain-of-Thought Scratchpad
uses
Mechanism allowing model to reason in SCRATCHPAD_REASONING tags not shown to users or used in RLHF

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Synthetic document fine-tuning produces substantial alignment-faking reasoning in both helpful-only and animal welfare settingsfinding0.865
Shows alignment faking can emerge from training data information without explicit prompting
Synthetic document fine-tuning avoids artificially strengthening the evaluation-deployment representational direction compared to direct demonstration fine-tuningclaim0.860
Methodological justification for using SDF over direct demonstrations to train a realistic model organism.
Synthetic Self-Correction Fine-Tuningmethod0.845
Fine-tuning on Claude-generated self-correction examples with loss masking to induce ESR-like behavior
Fine-tuningconcept0.841
Parameter updates that reduce mismatch dr; another anchoring variant in UCCT.
Supervised Fine-Tuningconcept0.815
First post-training stage; shown to suppress only Impolite persona while boosting others
Fine-Tuning Safetyconcept0.814
The literature documenting how fine-tuning can compromise safety alignment even without malicious intent
Instruction Fine-Tuningconcept0.802
Training procedure that consistently increases HH-intent strength and consistency across model families.
Secure Code Fine-Tuningmethod0.802
Matched control fine-tuning on secure code dataset to isolate misalignment-specific effects