2310.03714
#13
by
julien-c
HF staff
- opened
data.json
CHANGED
|
@@ -247,7 +247,7 @@
|
|
| 247 |
"abstract": "The ML community is rapidly exploring techniques for prompting language models (LMs) and for stacking them into pipelines that solve complex tasks. Unfortunately, existing LM pipelines are typically implemented using hard-coded \u201cprompt templates\u201d, i.e. lengthy strings discovered via trial and error. Toward a more systematic approach for developing and optimizing LM pipelines, we introduce DSPy, a programming model that abstracts LM pipelines as text transformation graphs, or imperative computational graphs where LMs are invoked through declarative modules. DSPy modules are parameterized, meaning they can learn (by creating and collecting demonstrations) how to apply compositions of prompting, finetuning, augmentation, and reasoning techniques. We design a compiler that will optimize any DSPy pipeline to maximize a given metric. We conduct two case studies, showing that succinct DSPy programs can express and optimize sophisticated LM pipelines that reason about math word problems, tackle multi-hop retrieval, answer complex questions, and control agent loops. Within minutes of compiling, DSPy can automatically produce prompt pipelines and finetune pipelines that outperform out-of-the-box few-shot prompting as well as expert-created demonstrations for GPT-3.5 and Llama2-13b-chat. On top of that, DSPy programs compiled to relatively small LMs like 770M parameter T5 and Llama2- 13b-chat are competitive with many approaches that rely on large and proprietary LMs like GPT-3.5 and on expert-written prompt chains.",
|
| 248 |
"type": "Spotlight Poster",
|
| 249 |
"OpenReview": "https://openreview.net/forum?id=sY5N0zY5Od",
|
| 250 |
-
"arxiv_id": "",
|
| 251 |
"GitHub": [],
|
| 252 |
"Space": [],
|
| 253 |
"Model": [],
|
|
|
|
| 247 |
"abstract": "The ML community is rapidly exploring techniques for prompting language models (LMs) and for stacking them into pipelines that solve complex tasks. Unfortunately, existing LM pipelines are typically implemented using hard-coded \u201cprompt templates\u201d, i.e. lengthy strings discovered via trial and error. Toward a more systematic approach for developing and optimizing LM pipelines, we introduce DSPy, a programming model that abstracts LM pipelines as text transformation graphs, or imperative computational graphs where LMs are invoked through declarative modules. DSPy modules are parameterized, meaning they can learn (by creating and collecting demonstrations) how to apply compositions of prompting, finetuning, augmentation, and reasoning techniques. We design a compiler that will optimize any DSPy pipeline to maximize a given metric. We conduct two case studies, showing that succinct DSPy programs can express and optimize sophisticated LM pipelines that reason about math word problems, tackle multi-hop retrieval, answer complex questions, and control agent loops. Within minutes of compiling, DSPy can automatically produce prompt pipelines and finetune pipelines that outperform out-of-the-box few-shot prompting as well as expert-created demonstrations for GPT-3.5 and Llama2-13b-chat. On top of that, DSPy programs compiled to relatively small LMs like 770M parameter T5 and Llama2- 13b-chat are competitive with many approaches that rely on large and proprietary LMs like GPT-3.5 and on expert-written prompt chains.",
|
| 248 |
"type": "Spotlight Poster",
|
| 249 |
"OpenReview": "https://openreview.net/forum?id=sY5N0zY5Od",
|
| 250 |
+
"arxiv_id": "2310.03714",
|
| 251 |
"GitHub": [],
|
| 252 |
"Space": [],
|
| 253 |
"Model": [],
|