contrastive-search

Description

Implementation of Contrastive Search, a decoding strategy that jointly optimizes model confidence and a degeneration penalty to produce fluent, coherent, and low-repetition text. At each step, the model considers the top-k candidate tokens and selects the one maximizing:

score(v) = (1 - alpha) * p(v | context) - alpha * max_cosine_similarity(h_v, H_context)

where alpha is the trade-off between confidence and the cosine-similarity-based penalty.

This strategy typically:

• Reduces repetition compared to greedy/beam search
• Preserves semantic coherence better than pure sampling

---

Base model

• Qwen/Qwen2.5-0.5B-Instruct (example)

---

Model compatibility

• Decoder and encoder-decoder transformer models for causal LM

---

Additional Arguments

• top_k (int): Number of candidate tokens to consider each step (e.g., 4)
• penalty_alpha (float): Weight of the degeneration penalty (e.g., 0.6)

Tips:

• Larger top_k explores more candidates but increases compute
• penalty_alpha in [0.3, 0.8] often works well; 0.0 reduces to greedy

---

Output Type changes

(none) — returns the same structure as standard transformers generation

---

Example usage

from transformers import AutoModelForCausalLM, AutoTokenizer, infer_device

device = infer_device()

model_id = "Qwen/Qwen2.5-0.5B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto").to(device)

inputs = tokenizer(["DeepMind Company is"], return_tensors="pt").to(device)

# Contrastive search
gen_out = model.generate(
    **inputs,
    custom_generate="contrastive_search",
    penalty_alpha=0.6,
    top_k=4,
    max_new_tokens=128,
    trust_remote_code=True,
)

print(tokenizer.batch_decode(gen_out, skip_special_tokens=True))