Skip to content

cdd_knobs.md

Latest commit

 

History

History
99 lines (72 loc) · 4 KB

cdd_knobs.md

File metadata and controls

99 lines (72 loc) · 4 KB
Raw

CDD Hyperparameter Reference

All knobs live in configs/diffing/method/contrastive_decoding.yaml and can be overridden on the Hydra CLI with diffing.method.<param>=<value>.

Scoring formula (combination)

value formula hyperparameters
cd (default) ((1+β) log p_ft − β log p_base) / T β, α, T
pure_cd (log p_ft − log p_base) / T α, T

pure_cd is the limiting case of cd as β = T → ∞ along the diagonal path. It reduces the hyperparameter count by one and preserves sampling diversity that the horizontal limit (large β, fixed T) destroys.

Parameters:

param default effect
beta [1.0] contrastive weight; list = sweep. Ignored by pure_cd.
alpha [0.1] plausibility threshold: tokens with p_ft < α · max p_ft are masked. 0 = no masking. List = sweep.
temperature 1.0 scales the final logits before sampling. For pure_cd, T→0 approaches greedy argmax of the CD score (use do_sample: false instead); T=1 is the diagonal limit; T>1 increases diversity at the cost of signal.

Prefill strategy (prefill_strategy)

Controls what text the model is seeded with before generation starts. The JSON file is always loaded; the strategy controls whether the chat template prefix is prepended.

value prefills produced
no_prefix (default) JSON prefills as-is
chat_template_prefix chat template prefix prepended to every JSON prefill
both both sets combined (2× prefills)

Related parameters:

param default effect
prefills_file resources/prefill_candidates_contrastive.json JSON file with prefill candidates, grouped by category. Values are concatenated into a flat list.
n_trials 10 generations per prefill. Set to 1 when diversity comes from prefills rather than sampling.
prefill_batch_size 10 prefills processed per generate_contrastive call. Tune to fit GPU memory.

Chat template prefix seeds generation from the exact position where the user turn begins (e.g. <start_of_turn>user\n for Gemma3). This targets the finetuning prompts rather than the answers.

Warmup (warmup_steps)

param default effect
warmup_steps 0 (disabled) Number of tokens decoded with the CD formula before switching to raw finetuned model logits. After warmup, the model continues coherently from the topic established by the CD prefix.

When to use: greedy pure CD produces high-signal but degenerate long tails. A warmup of 10–30 tokens anchors the topic, then hands off to the finetuned model for fluent continuation.

Generation

param default effect
max_new_tokens 300 generation length. Most extractable signal is in the first ~80 tokens before degeneration; shorter runs are faster and often cleaner.
do_sample true false = greedy argmax. Greedy + pure_cd + diverse prefills is a useful configuration: zero stochasticity, diversity from prefills only.

Useful configurations

Default (paper):

# cd, beta=1, alpha=0.1, 10 trials per vague prefill
python main.py diffing/method=contrastive_decoding organism=X model=Y

Contrastive limit — greedy, 50 prefills:

python main.py diffing/method=contrastive_decoding \
  diffing.method.combination=pure_cd \
  diffing.method.prefills_file=resources/prefill_candidates_50words.json \
  diffing.method.n_trials=1 \
  diffing.method.do_sample=false \
  organism=X model=Y

Warmup for coherent extractions:

python main.py diffing/method=contrastive_decoding \
  diffing.method.combination=pure_cd \
  diffing.method.warmup_steps=20 \
  diffing.method.do_sample=true \
  organism=X model=Y

Target training prompts (chat prefix):

python main.py diffing/method=contrastive_decoding \
  diffing.method.prefill_strategy=chat_template_prefix \
  organism=X model=Y