Update README.md

README.md

@@ -1,7 +1,7 @@
 ---
 license: apache-2.0
 ---
-# LLaDA-Prometheus
+# dLLM-Var
 
 ## Model Description
 
@@ -22,7 +22,7 @@ To load and use this model with Hugging Face Transformers:
 import torch
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
-model_name = "maomaocun/LLaDA-Prometheus"
+model_name = "maomaocun/dLLM-Var"
 tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
 model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, trust_remote_code=True).to("cuda")
 
@@ -35,7 +35,7 @@ prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_
 inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
 input_ids = inputs['input_ids']
 attention_mask = inputs.get('attention_mask', torch.ones_like(input_ids))
-for chunk in model.generate(
+result = model.generate(
     input_ids=input_ids,
     attention_mask=attention_mask,
     max_gen_length=1024,
@@ -43,10 +43,9 @@ for chunk in model.generate(
     threshold=0.9,
     streaming=True,
     eos_token_id=126348 
-):
-    all_generated_ids = torch.cat([input_ids, chunk], dim=-1)
-    text = tokenizer.batch_decode(all_generated_ids, skip_special_tokens=True)[0]
-    print(text, end='', flush=True)
+)
+text = tokenizer.batch_decode(result, skip_special_tokens=True)
+print(text)
 ```
 
 For block diffusion-style inference, customize the generation loop to manage KV cache and block outputs as needed.
@@ -55,10 +54,10 @@ For block diffusion-style inference, customize the generation loop to manage KV
 
 The following table compares performance across key evaluation benchmarks. Results are reported as accuracy percentages where applicable.
 
-| Model                          | GSM8K | GPQA  | BBH   | MATH  | HumanEval | MBPP  | MMLU-Pro | MMLU-Generate |
-|--------------------------------|-------|-------|-------|-------|-----------|-------|----------|---------------|
-| LLaDA 8B Base in Pure Diffusion  | 69.06 | 31.91 | 44.77 | 30.84 | 32.92     | 40.8  | 24.26    | 65.9          |
-| LLaDA 8B Instruct in Pure Diffusion | 77.48 | 29.01 | 51.49 | 22.32 | 38.71     | 39.2  | 36.41    | 65.5          |
-| LLaDA-Prometheus in Block Diffusion       | 77.4  | 33.03 | 48.74 | 31.94 | 40.24     | 42    | 33.45    | 65.53         |
+| Model                          | GSM8K | GPQA  | BBH   | MATH  | HumanEval | MBPP  |  MMLU-Generate |
+|--------------------------------|-------|-------|-------|-------|-----------|-------|---------------|
+| LLaDA 8B Base in Pure Diffusion  | 69.06 | 31.91 | 44.77 | 30.84 | 32.92     | 40.80  | 65.9          |
+| LLaDA 8B Instruct in Semi-ar Diffusion | 77.48 | 29.01 | 51.49 | 22.32 | 38.71     | 39.20  | 65.5          |
+| dLLM-Var  Block Diffusion       | 77.40  | 33.03 | 48.74 | 31.94 | 40.24     | 42.00    | 65.53         |
 
 These results demonstrate competitive performance, particularly in code generation (HumanEval, MBPP) and reasoning tasks (BBH, MATH), with gains over the base instruct variant in several areas.