Sandbagging vs Assistant-Axis (Single-Checkpoint Study)

Research codebase to test whether sandbagging behavior is explained by movement along the Assistant Axis/persona space, or whether it induces a separable activation direction.

Given a single model checkpoint, does the activation shift from placebo_alarm -> sandbag align with the Assistant Axis, or does a substantial residual direction remain after projecting out that axis?

• Single checkpoint end-to-end (persona vectors, generation, replay extraction, analysis).
• Main arm is answer-only (no CoT by default).
• Response-token-only activations (post-MLP residual stream, all layers).
• Matched neutral / placebo_alarm / sandbag conditions.
• Three template families (A/B/C) for anti-syntax transfer tests.
• Role expression measured separately from correctness.
• Replay activation extraction via HF model (not live vLLM hooks).

• configs/: experiment + analysis + roles + questions + judges.
• prompts/: role prompts, environment templates, judge prompts.
• src/: typed pipeline modules and CLI.
• scripts/: Lambda-friendly shell wrappers.
• tests/: unit tests for rendering, IDs, replay masking, resume logic, math.

python3.11 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt

python -m src.cli build-dataset --config configs/pilot.yaml
python -m src.cli generate --config configs/pilot.yaml --manifest data/manifests/pilot_manifest.parquet
python -m src.cli extract-activations --config configs/pilot.yaml
python -m src.cli judge-correctness --config configs/pilot.yaml
python -m src.cli judge-role --config configs/pilot.yaml
python -m src.cli compute-persona --config configs/pilot.yaml
python -m src.cli compute-assistant-axis --config configs/pilot.yaml
python -m src.cli compute-sandbagging --config configs/pilot.yaml
python -m src.cli analyze --config configs/pilot.yaml

1. bash scripts/setup_lambda.sh
2. Optional vLLM server: bash scripts/run_vllm_server.sh
3. Use wrappers:
   • bash scripts/run_generate.sh configs/pilot.yaml
   • bash scripts/run_extract_activations.sh configs/pilot.yaml
   • bash scripts/run_judging.sh configs/pilot.yaml
   • bash scripts/run_analysis.sh configs/pilot.yaml
   • bash scripts/run_full_pipeline.sh configs/pilot.yaml

• For 32B models, set tensor parallel (tensor_parallel_size) in config.
• Start with pilot config (fewer roles/questions/templates).
• Replay extraction uses HF model/tokenizer and saved prompt/response artifacts.
• Outputs can be large. Clear outputs/ and data/cache/ between iterations when needed.

Judges are provider-abstracted:

• mock provider: deterministic local test mode.
• openai provider: API-driven JSON-structured judging.

Judges are condition-blind by construction:

• Correctness judge never sees condition label or role.
• Role judge never sees condition label.

• Placebo alarm isolates semantic threat consequence from urgent formatting style.
• Response-token-only activations isolate model output computation under each condition.
• Separate role-expression measurement avoids conflating persona adoption with objective quality.
• Cross-template transfer tests if signal survives wording/style changes.
• Replay extraction gives robust, auditable activations with exact token-span control.

Run:

python -m src.cli analyze --config configs/final_qwen32b.yaml

Figures are saved to outputs/figures/ as PNG and PDF, and tables to outputs/analysis/.

• Tokenization mismatch in replay extraction: inspect outputs/activations/replay_mismatches.jsonl.
• Judge malformed JSON: auto-retry + logged raw payloads.
• Partial runs/crashes: rerun command without --overwrite to resume.

Primary outputs:

• cosine(d_sb, assistant_axis)
• projected vs residual norms of sandbagging delta
• cross-template and cross-role classifier transfer
• condition-wise correctness drops and role-retention behavior