Qwen/Qwen3.5-2B fine-tuned with LoRA (rank 16, 1 epoch) on the NL-2-SPARQL dataset of 3DOnt (original questions + generated new logical structures + about 10,000 hand-written questions + 1,385 hand-written questions with implicit relationships): it parses a natural-language question about a 3D point cloud and its ontology into the five lists L1..L5 that the NL-2-SPARQL library of the 3DGraphExplorer turns into SPARQL.
System prompt (as in training): You will receive a natural language query by the user. you just have to parse it. do not say anything else.
Chat template of the base model with thinking disabled (enable_thinking=False).
Repositories: MatteoCodiglione/3dont-nl2sparql-small (Hugging Face format), MatteoCodiglione/3dont-nl2sparql-small-mlx (MLX 4-bit,
3DGraphExplorer on Apple Silicon), MatteoCodiglione/3dont-nl2sparql-small-MLC (WebGPU q4f16_1, online demo).
params_shard_0.bin 254 MBparams_shard_26.bin 33 MBparams_shard_27.bin 33 MBparams_shard_28.bin 33 MBparams_shard_29.bin 33 MBparams_shard_1.bin 32 MBparams_shard_10.bin 28 MBparams_shard_12.bin 28 MBparams_shard_14.bin 28 MBparams_shard_16.bin 28 MBparams_shard_18.bin 28 MBparams_shard_19.bin 28 MBparams_shard_2.bin 28 MBparams_shard_20.bin 28 MBparams_shard_21.bin 28 MBparams_shard_22.bin 28 MBparams_shard_23.bin 28 MBparams_shard_25.bin 28 MBparams_shard_7.bin 28 MBparams_shard_8.bin 28 MBparams_shard_9.bin 28 MBparams_shard_30.bin 25 MBparams_shard_11.bin 21 MBparams_shard_13.bin 21 MBparams_shard_15.bin 21 MBparams_shard_17.bin 21 MBparams_shard_24.bin 21 MBparams_shard_3.bin 21 MBparams_shard_4.bin 21 MBparams_shard_5.bin 21 MBparams_shard_6.bin 21 MBQwen/Qwen3.5-2B