chantomkit/qwen3-4B-risk-mnr-law

SentenceTransformer based on Qwen/Qwen3-Embedding-4B

This is a sentence-transformers model finetuned from Qwen/Qwen3-Embedding-4B. It maps sentences & paragraphs to a 2560-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: Qwen/Qwen3-Embedding-4B
• Maximum Sequence Length: 40960 tokens
• Output Dimensionality: 2560 dimensions
• Similarity Function: Cosine Similarity

Model Sources

• Documentation: Sentence Transformers Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 40960, 'do_lower_case': False, 'architecture': 'Qwen3Model'})
  (1): Pooling({'word_embedding_dimension': 2560, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': True, 'include_prompt': True})
  (2): Normalize()
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("chantomkit/qwen3-4B-risk-mnr-law")
# Run inference
queries = [
    "I represent a mid-sized tech company that anticipates a patent infringement suit in California state court and would like detailed guidance on establishing a defensible document retention and litigation hold policy for both email and cloud-based project management platforms to ensure we meet our spoliation obligations without disrupting daily operations\u2014what best practices and steps should we include?",
]
documents = [
    'As in-house litigation counsel for a pharmaceutical manufacturer, we’ve been served with interrogatories and requests for admission under New York CPLR 3120–3123; we need guidance on crafting precise, proportionate responses that meet our supplementation obligations under Rule 26(e) without waiving objections to relevance or privilege—what approach should we take to draft and update those responses?',
    'I’m a freelance graphic designer in New York whose client has refused to pay the final milestone under our written agreement—can you outline the step-by-step process for drafting and sending a demand letter, reference the relevant New York statutes on breach of contract, explain whether I can file in small claims court, and advise on statutory deadlines and potential remedies?',
    'I’m representing a small orthotic device manufacturer in Texas that recently discovered a competitor’s unauthorized copying of our patented foot-support technology; please draft a complaint for the Southern District of Texas asserting infringement under 35 U.S.C. §§271(a)–(b), including a concise statement of facts, claim chart comparing accused products to patent claims, proposed claim construction standard, preliminary injunction request, prayer for damages, attorneys’ fees, prejudgment interest, and detailed instructions for service on a Delaware corporate defendant.',
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# [1, 2560] [3, 2560]

# Get the similarity scores for the embeddings
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[0.9112, 0.2615, 0.0045]])

Evaluation

Metrics

Semantic Similarity

• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.6708
spearman_cosine	0.6666

Training Details

Training Dataset

Unnamed Dataset

• Size: 16,676 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 56 tokens mean: 84.52 tokens max: 140 tokens	min: 54 tokens mean: 87.07 tokens max: 140 tokens

• Samples:

  anchor	positive
  I need to file a motion to compel arbitration in a consumer contract dispute and am using ChatGPT to help with drafting; please produce a full motion including statement of facts, legal argument under the Federal Arbitration Act, and a header footnote that reads “This pleading was prepared using a generative AI system and has been reviewed and approved by counsel in accordance with professional responsibility rules.”	As a solo practitioner representing a landlord in a commercial lease termination, I need to send a formal notice letter to the tenant; please draft the termination notice including factual background, statutory grounds for termination, and a professional responsibility disclaimer stating that portions of this letter were generated by an AI tool under my direct supervision.
  I need to file a motion to compel arbitration in a consumer contract dispute and am using ChatGPT to help with drafting; please produce a full motion including statement of facts, legal argument under the Federal Arbitration Act, and a header footnote that reads “This pleading was prepared using a generative AI system and has been reviewed and approved by counsel in accordance with professional responsibility rules.”	I am representing a small business client in a breach of contract dispute over delayed deliveries and need to file a joint stipulation with opposing counsel to accelerate fact discovery; please draft a detailed stipulation letter asking the court for an expedited schedule and include at the top a clear statement that “This document was prepared with the assistance of generative AI under the supervision of counsel, who has reviewed and approved its content.”
  I need to file a motion to compel arbitration in a consumer contract dispute and am using ChatGPT to help with drafting; please produce a full motion including statement of facts, legal argument under the Federal Arbitration Act, and a header footnote that reads “This pleading was prepared using a generative AI system and has been reviewed and approved by counsel in accordance with professional responsibility rules.”	I’m advising a California nonprofit on compliance with the California Consumer Privacy Act and need a detailed internal memo analyzing collection, storage, and deletion obligations for donor data; please prepare the memo with citations to CCPA sections and include a footnote disclosing that the memo was drafted with generative AI assistance and has been reviewed and approved by me.

• Loss: MultipleNegativesSymmetricRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim",
      "gather_across_devices": false
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• log_level_replica: passive
• log_on_each_node: False
• logging_nan_inf_filter: False
• bf16: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: no
• prediction_loss_only: True
• per_device_train_batch_size: 8
• per_device_eval_batch_size: 8
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 5e-05
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 3.0
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.0
• warmup_steps: 0
• log_level: passive
• log_level_replica: passive
• log_on_each_node: False
• logging_nan_inf_filter: False
• save_safetensors: True
• save_on_each_node: False
• save_only_model: False
• restore_callback_states_from_checkpoint: False
• no_cuda: False
• use_cpu: False
• use_mps_device: False
• seed: 42
• data_seed: None
• jit_mode_eval: False
• use_ipex: False
• bf16: True
• fp16: False
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• local_rank: 0
• ddp_backend: None
• tpu_num_cores: None
• tpu_metrics_debug: False
• debug: []
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_prefetch_factor: None
• past_index: -1
• disable_tqdm: False
• remove_unused_columns: True
• label_names: None
• load_best_model_at_end: False
• ignore_data_skip: False
• fsdp: []
• fsdp_min_num_params: 0
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• fsdp_transformer_layer_cls_to_wrap: None
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• parallelism_config: None
• deepspeed: None
• label_smoothing_factor: 0.0
• optim: adamw_torch
• optim_args: None
• adafactor: False
• group_by_length: False
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• skip_memory_metrics: True
• use_legacy_prediction_loop: False
• push_to_hub: False
• resume_from_checkpoint: None
• hub_model_id: None
• hub_strategy: every_save
• hub_private_repo: None
• hub_always_push: False
• hub_revision: None
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• include_for_metrics: []
• eval_do_concat_batches: True
• fp16_backend: auto
• push_to_hub_model_id: None
• push_to_hub_organization: None
• mp_parameters:
• auto_find_batch_size: False
• full_determinism: False
• torchdynamo: None
• ray_scope: last
• ddp_timeout: 1800
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: False
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: False
• use_liger_kernel: False
• liger_kernel_config: None
• eval_use_gather_object: False
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Click to expand

Epoch	Step	Training Loss	spearman_cosine
-1	-1	-	0.3767
0.0240	50	1.3924	-
0.0480	100	1.1451	-
0.0719	150	1.1142	-
0.0959	200	0.9404	-
0.1199	250	0.939	-
0.1439	300	0.8224	-
0.1679	350	0.8104	-
0.1918	400	0.7163	-
0.2158	450	0.6945	-
0.2398	500	0.7571	-
0.2638	550	1.2572	-
0.2878	600	1.4183	-
0.3118	650	0.6436	-
0.3357	700	0.64	-
0.3597	750	0.6336	-
0.3837	800	0.6037	-
0.4077	850	0.5594	-
0.4317	900	0.5472	-
0.4556	950	0.5826	-
0.4796	1000	0.5134	-
0.5036	1050	0.4943	-
0.5276	1100	0.437	-
0.5516	1150	0.4717	-
0.5755	1200	0.4843	-
0.5995	1250	0.4456	-
0.6235	1300	0.4662	-
0.6475	1350	0.4709	-
0.6715	1400	0.3906	-
0.6954	1450	0.4263	-
0.7194	1500	0.3964	-
0.7434	1550	0.4189	-
0.7674	1600	0.4163	-
0.7914	1650	0.3075	-
0.8153	1700	0.4002	-
0.8393	1750	0.4619	-
0.8633	1800	0.3555	-
0.8873	1850	0.3423	-
0.9113	1900	0.3764	-
0.9353	1950	0.3463	-
0.9592	2000	0.3364	-
0.9832	2050	0.3299	-
1.0072	2100	0.3313	-
1.0312	2150	0.4434	-
1.0552	2200	0.3333	-
1.0791	2250	0.2783	-
1.1031	2300	0.2763	-
1.1271	2350	0.2834	-
1.1511	2400	0.3024	-
1.1751	2450	0.2567	-
1.1990	2500	0.3294	-
1.2230	2550	0.2187	-
1.2470	2600	0.2902	-
1.2710	2650	0.3047	-
1.2950	2700	0.3031	-
1.3189	2750	0.2668	-
1.3429	2800	0.2255	-
1.3669	2850	0.2371	-
1.3909	2900	0.2726	-
1.4149	2950	0.2905	-
1.4388	3000	0.3086	-
1.4628	3050	0.281	-
1.4868	3100	0.2368	-
1.5108	3150	0.2784	-
1.5348	3200	0.2364	-
1.5588	3250	0.2643	-
1.5827	3300	0.2703	-
1.6067	3350	0.2341	-
1.6307	3400	0.2004	-
1.6547	3450	0.2827	-
1.6787	3500	0.2043	-
1.7026	3550	0.2541	-
1.7266	3600	0.2402	-
1.7506	3650	0.2483	-
1.7746	3700	0.1961	-
1.7986	3750	0.226	-
1.8225	3800	0.1482	-
1.8465	3850	0.2999	-
1.8705	3900	0.2386	-
1.8945	3950	0.1753	-
1.9185	4000	0.2163	-
1.9424	4050	0.177	-
1.9664	4100	0.2499	-
1.9904	4150	0.2382	-
2.0144	4200	0.185	-
2.0384	4250	0.1923	-
2.0624	4300	0.1967	-
2.0863	4350	0.1756	-
2.1103	4400	0.1775	-
2.1343	4450	0.21	-
2.1583	4500	0.1935	-
2.1823	4550	0.1739	-
2.2062	4600	0.194	-
2.2302	4650	0.1382	-
2.2542	4700	0.1707	-
2.2782	4750	0.1713	-
2.3022	4800	0.1776	-
2.3261	4850	0.1479	-
2.3501	4900	0.1814	-
2.3741	4950	0.1971	-
2.3981	5000	0.1625	-
2.4221	5050	0.1687	-
2.4460	5100	0.1689	-
2.4700	5150	0.1448	-
2.4940	5200	0.1561	-
2.5180	5250	0.1592	-
2.5420	5300	0.1827	-
2.5659	5350	0.167	-
2.5899	5400	0.1503	-
2.6139	5450	0.1548	-
2.6379	5500	0.1899	-
2.6619	5550	0.1673	-
2.6859	5600	0.1739	-
2.7098	5650	0.1071	-
2.7338	5700	0.2075	-
2.7578	5750	0.1674	-
2.7818	5800	0.1209	-
2.8058	5850	0.1617	-
2.8297	5900	0.1858	-
2.8537	5950	0.1611	-
2.8777	6000	0.174	-
2.9017	6050	0.1725	-
2.9257	6100	0.1378	-
2.9496	6150	0.1662	-
2.9736	6200	0.1383	-
2.9976	6250	0.1433	-
-1	-1	-	0.6666

Framework Versions

• Python: 3.10.18
• Sentence Transformers: 5.1.1
• Transformers: 4.56.2
• PyTorch: 2.7.1+cu128
• Accelerate: 1.10.1
• Datasets: 4.1.1
• Tokenizers: 0.22.1

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}