Edit model card

GraphWiz

Project Page: https://graph-wiz.github.io/

Paper: https://arxiv.org/abs/2402.16029.pdf

Code: https://github.com/nuochenpku/Graph-Reasoning-LLM

GraphWiz is a powerful instruction-following LLM that can map textural descriptions of graphs and structures, and then solve different graph problems explicitly in natural language.

Training strategies include two stages: Mixed-task Training and DPO Alignment.

Results

Models Cycle Connect Bipartite Topology Shortest Triangle Flow Hamilton Subgraph Average
In-Context Learning
GPT-4 (zero-shot) 38.75 17.00 65.25 5.00 9.25 5.75 3.25 59.25 45.50 27.67
GhatGPT (2-shot) 51.25 43.75 70.75 4.50 3.50 17.25 8.50 54.25 43.00 32.97
GPT-4 (2-shot) 52.50 62.75 74.25 25.25 18.25 31.00 7.75 {75.75} 46.75 43.81
Mistral-7B
Naive SFT 73.75 83.50 78.50 1.00 23.00 47.00 28.75 31.75 41.25 46.56
GraphWiz 92.00 89.50 72.00 19.00 31.25 38.75 29.25 26.50 85.50 53.75
GraphWiz-DPO 85.50 79.50 85.50 85.25 12.50 29.00 35.50 62.75 48.50 58.22
LLaMA 2-7B
Naive SFT 73.75 83.50 41.25 4.00 9.50 30.00 16.50 69.00 75.45 44.81
GraphWiz 91.50 87.00 74.00 18.00 28.00 38.25 24.50 52.25 82.25 55.08
GraphWiz-DPO 89.00 82.50 84.75 46.75 24.00 52.75 43.50 81.50 77.25 65.00
LLaMA 2-13B
Naive SFT 73.75 83.75 59.00 0.50 11.75 34.75 24.25 59.75 54.75 44.69
GraphWiz 94.75 87.00 78.00 28.00 27.75 36.00 24.50 59.00 81.50 57.39
GraphWiz-DPO 87.50 88.50 88.25 72.75 22.00 48.75 43.75 46.50 77.00 63.89

Examples

G-Q: Determine whether or not there is a cycle in an undirected graph. In an undirected graph..,the
 nodes are numbered from 0 to 88, and the edges are: (0, 73) (0, 51) (0, 10) (0, 63) (0, 28) (1, 62) (1, 57) (1, 84) (1, 61) (1, 5)
 (1, 24) (2, 84) (2, 3) (2, 66) (2, 68) (2, 17) (2, 35) (2, 34) (2, 15) (3, 39) (3, 52) (3, 16) (3, 15) (3, 8) (4, 69) (4, 85)
 (4, 36) (4, 72) (5, 44) (6, 77) (6, 7) (7, 85) (8, 64) (8, 23) (8, 28) (9, 34) (9, 31) (9, 61) (9, 28) (10, 26) (11, 37) (11, 39)
 (11, 19) (11, 64) (13, 73) (13, 61) (13, 80) (13, 85) (14, 86) (14, 59) (14, 32) (14, 58) (14, 85) (14, 66) (15, 43) (15, 48) (15, 73)
 (15, 19) (15, 47) (15, 68) (16, 46) (16, 60) (16, 84) (17, 44) (17, 72) (17, 36) (17, 37) (17, 61) (18, 20) (18, 24) (18, 22) (18, 41)
 (19, 45) (19, 83) (20, 25) (20, 29) (21, 38) (21, 64) (21, 24) (21, 22) (21, 34) (22, 23) (23, 34) (23, 30) (23, 83) (24, 47) (24, 50)
 (25, 59) (25, 42) (25, 70) (25, 72) (26, 45) (26, 30) (26, 87) (26, 80) (26, 50) (27, 77) (27, 58) (27, 60) (27, 29) (28, 36) (28, 59)
 (28, 88) (29, 31) (29, 33) (29, 56) (29, 55) (29, 78) (29, 73) (29, 82) (30, 34) (31, 58) (31, 70) (32, 78) (33, 73) (33, 68) (34, 48)
(34, 88) (34, 51) (34, 66) (34, 54) (35, 43) (35, 56) (36, 44) (36, 39) (37, 57) (37, 71) (38, 47) (38, 58) (38, 88) (39, 74) (39, 48)
 (39, 45) (39, 42) (40, 60) (40, 59) (40, 71) (40, 76) (41, 79) (42, 60) (42, 62) (42, 72) (43, 57) (44, 47) (44, 74) (44, 75) (44, 82)
 (44, 70) (45, 68) (45, 67) (45, 71) (46, 84) (47, 56) (48, 75) (49, 71) (49, 58) (50, 76) (51, 55) (51, 87) (51, 79) (52, 78) (56, 69)
(56, 83) (56, 79) (57, 76) (57, 70) (57, 72) (57, 78) (59, 70) (60, 80) (60, 76) (60, 77) (60, 75) (61, 74) (61, 77) (62, 84) (62, 66)
 (66, 69) (68, 88) (70, 88) (70, 76) (71, 81) (71, 72) (71, 86) (72, 75) (73, 82) (73, 83) (74, 85). Is there a cycle in this graph?


GraphWiz: 
There are many possible ways to find a cycle in this graph, but one example is: Starting from node 0, we can go to node 73 (via edge 0-73), then to node 82 (via edge 73-82),
then to node 44 (via edge 82-44), then to node 17 (via edge 44-17), then to node 36 (via edge 17-36), then to node 28 (via edge 36-28),
and back to node 0 (via edge 28-0). This forms a cycle [0-73-82-44-17-36-28-0] without revisiting any edge. So, there is a cycle in this graph.
Downloads last month
16
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train GraphWiz/LLaMA2-13B-RFT