mmardani commited on
Commit
72abdaf
·
verified ·
1 Parent(s): 7eb33c7

Point evaluation results at the technical report; document the per-task prompting convention

Browse files
Files changed (1) hide show
  1. README.md +11 -2
README.md CHANGED
@@ -165,10 +165,19 @@ Below roughly 16 steps the pretrained models lose quality noticeably; for few-st
165
  prefer the distilled `-distill-*` variants, which are trained for a fixed step count and need
166
  neither guidance nor score-temperature tuning.
167
 
 
 
 
 
 
 
 
 
168
  ## Evaluation Results
169
 
170
- Evaluation on the public benchmarks listed above is being finalized; results
171
- will be published with the release.
 
172
 
173
  ## Training, Testing, and Evaluation Datasets
174
 
 
165
  prefer the distilled `-distill-*` variants, which are trained for a fixed step count and need
166
  neither guidance nor score-temperature tuning.
167
 
168
+ ### Prompting convention
169
+
170
+ Use the chat template for conversational and reasoning tasks (see the example
171
+ above). For code-completion benchmarks such as HumanEval and MBPP, pass the raw
172
+ prompt **without** the chat template: those tasks hand the model a function
173
+ signature to continue, and wrapping them in a chat turn makes the model answer
174
+ conversationally instead of completing the code.
175
+
176
  ## Evaluation Results
177
 
178
+ Benchmark results for this model family, averaged over multiple generation
179
+ seeds, are reported in the technical report:
180
+ [Large Language Continuous Diffusion Models](https://arxiv.org/abs/2610.02665).
181
 
182
  ## Training, Testing, and Evaluation Datasets
183