How you evaluate reasoning models like QwQ-32B, since the response time and token length is very long?

[txchen-USTC]

Any adjustments to the hyperparameters in pred.py?

[bys0318]

We use Temperature=0.6, TopP=0.95, MinP=0, max_new_tokens=30000, max_input_len=100000. Remember to enable YaRN whilst evaluating.

[txchen-USTC]

I wonder whether you need to add an additional hyperpamameter "timeout" to the following place during evaluation:

completion = client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": prompt}],
temperature=temperature,
max_tokens=max_new_tokens,
timeout=40000 # or larger?
)

and for enabling YARN, is the following config of QwQ-32B right for evaluation?

{
"architectures": [
"Qwen2ForCausalLM"
],
"attention_dropout": 0.0,
"bos_token_id": 151643,
"eos_token_id": 151645,
"hidden_act": "silu",
"hidden_size": 5120,
"initializer_range": 0.02,
"intermediate_size": 27648,
"max_position_embeddings": 40960,
"max_window_layers": 64,
"model_type": "qwen2",
"num_attention_heads": 40,
"num_hidden_layers": 64,
"num_key_value_heads": 8,
"rms_norm_eps": 1e-05,
"rope_theta": 1000000.0,
"rope_scaling": {
"factor": 4.0,
"original_max_position_embeddings": 32768,
"type": "yarn"
},
"sliding_window": 131072,
"tie_word_embeddings": false,
"torch_dtype": "bfloat16",
"transformers_version": "4.43.1",
"use_cache": true,
"use_sliding_window": false,
"vocab_size": 152064
}

[bys0318]

I set timeout=3600. Your YaRN configuration is correct.

[txchen-USTC]

Thanks to your reply!

[txchen-USTC]

Have you evaluated QwQ-32B on Longbench v1? If so, are ther any adjustments to the hyperparameters in pred.py?