What it is. An extractive question-answering model. Given a question and a short typed answer, it returns the phrase in the answer that answers the question, or nothing. It cannot generate text. It is the reading part of rai, a tool that says only whether a description is complete under a notion of completeness written down in advance. Every decision after the reading is made in plain code. Trained from. deepset/minilm-uncased-squad2 (CC-BY-4.0), deepset's fine-tune of Microsoft's MiniLM-L12-H384-uncased (MIT) on SQuAD 2.0. Same architecture, 33M parameters, nothing added. Credit to deepset and Microsoft. 0.3.7 trains on 0.3.6's rows and 5,800 more written in Claude Code. In 0.3.6's rows…
Open-weight model · Question answering
minilm-uncased-squad2
by Deepset deepset/minilm-uncased-squad2
minilm-uncased-squad2 is an open-weight model for question answering from Deepset, released under Creative Commons Attribution 4.0. It has 33M parameters and a 512-token context. At 16-bit it needs about 0.1 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 21k downloads a month.
Haystack is an AI orchestration framework to build customizable, production-ready LLM applications. You can use this model in Haystack to do extractive question answering on documents.
Parameters33M
Context512
Weights399.8 MB
Licensecc-by-4.0
AccessOpen weights
Monthly Downloads21k
Runs On
What it takes to serve minilm-uncased-squad2 (33M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 0.1 GB | 0.1 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 0.0 GB | 0.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.0 GB | 0.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 10, 2026.
minilm-uncased-squad2 on every accelerator the SAVRN Index prices, at every precision
Model Card
By Deepset, published under cc-by-4.0, revision 934656cdda79.
MiniLM-L12-H384-uncased for Extractive QA
Overview
Language model: microsoft/MiniLM-L12-H384-uncased
Language: English
Downstream-task: Extractive QA
Training data: SQuAD 2.0
Eval data: SQuAD 2.0
Code: See an example extractive QA pipeline built with Haystack
Infrastructure: 1x Tesla v100
Hyperparameters
seed=42
batch_size = 12
n_epochs = 4
base_LM_model = "microsoft/MiniLM-L12-H384-uncased"
max_seq_len = 384
learning_rate = 4e-5
lr_schedule = LinearWarmup
warmup_proportion = 0.2
doc_stride=128
max_query_length=64
grad_acc_steps=4
Usage
In Haystack
Haystack is an AI orchestration framework to build customizable, production-ready LLM applications. You can use this model in Haystack to do extractive question answering on documents. To load and run the model with Haystack:
Configuration
- Architecture
- BertForQuestionAnswering
- Context length (tokens)
- 512
- Layers
- 12
- Hidden size
- 384
- Feed-forward size
- 1,536
- Attention heads
- 12
- Vocabulary size
- 30,522
- Model type
- bert
Identity and Version
- Repository
- deepset/minilm-uncased-squad2
- Publisher
- Deepset
- Task
- Question answering
- Modality
- Text
- Library
- transformers
- Parameters
- 33M parameters
- Languages
- en
- Revision
- 934656cdda79824eabf503ed56e15c01ddbdbe3f
- First published
- 2022-03-02
- Last updated
- 2024-09-26
Files and Weights
10 files, 400.1 MB in total. The weights are 4 files totalling 399.8 MB in bin, msgpack, safetensors.
Weights4 files · 399.8 MB
Configuration2 files · 589 B
Tokenizer2 files · 231.6 KB
Documentation1 file · 6.1 KB
Repository1 file · 445 B
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| flax_model.msgpack | Weights | 132.9 MB | bcd4d6e16c9e |
| model.safetensors | Weights | 133.5 MB | 240478c7251d |
| pytorch_model.bin | Weights | 133.5 MB | d8582ee8eec6 |
| training_args.bin | Weights | 1.5 KB | 1ce3a6d0039a |
| config.json | Configuration | 477 B | — |
| special_tokens_map.json | Configuration | 112 B | — |
| README.md | Documentation | 6.1 KB | — |
| .gitattributes | Repository | 445 B | — |
| tokenizer_config.json | Tokenizer | 107 B | — |
| vocab.txt | Tokenizer | 231.5 KB | — |
License and Download
- License
- cc-by-4.0
- Access
- Open weights, no gate
- Download size
- 399.8 MB
Download from Deepset
Released by Deepset through its official repository on Hugging Face. Read the license.
Built From
- Trained on (disclosed) squad_v2
Evaluations
Each result is shown as reported, with the conditions its reporter stated. None is a SAVRN measurement. A comparison lines two results up only when their configuration, unit and setup are all stated and identical.
| Benchmark | Conditions | Result | Reported by | Revision | Date |
|---|---|---|---|---|---|
| squad_v2 | Configuration squad_v2Task Question AnsweringMetric Exact MatchComparison conditions not established | 76.1921 | deepset Publisher reported |
Evaluated revision not stated | — |
| squad_v2 | Configuration squad_v2Task Question AnsweringMetric F1Comparison conditions not established | 79.5483 | deepset Publisher reported |
Evaluated revision not stated | — |
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 399.8 MB |
| 16-bit | 0.1 GB |
| 8-bit | 0.0 GB |
| 4-bit | 0.0 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Built on This Model
- Derived fromrai
Questions About minilm-uncased-squad2
How much GPU memory does minilm-uncased-squad2 need?
About 0.1 GB at 16-bit and 0 GB at 4-bit: the weights (33M parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run minilm-uncased-squad2 on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use minilm-uncased-squad2 commercially?
Yes. minilm-uncased-squad2 is released under Creative Commons Attribution 4.0. CC BY 4.0 permits sharing and adapting the work, including commercially, provided the creator is credited and changes are indicated.
What is minilm-uncased-squad2's context length?
512 tokens, from the maximum position embeddings in its published configuration.
Similar Models
This modelcard aims to be a base template for new models. It has been generated using this raw template. Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).
Haystack version 1.x distillation feature was used for training. deepset/bert-large-uncased-whole-word-masking-squad2 was used as the teacher model. Haystack is an AI orchestration framework to build customizable, production-ready LLM applications. You can use this model in Haystack to do extractive question answering on documents. To load and run the model with Haystack: For a complete example with an extractive question answering pipeline that scales over many documents, check out the corresponding Haystack tutorial. - Timo Möller: timo.moeller [at] deepset.ai deepset is the company behind the production-ready open-source AI framework Haystack. We also have a Discord community open to…
MobileBERT is a thin version of BERTLARGE, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks. This model was fine-tuned from the HuggingFace checkpoint google/mobilebert-uncased on SQuAD2.0. CPU: Intel(R) Core(TM) i7-6800K CPU @ 3.40GHz Memory: 32 GiB GPUs: 2 GeForce GTX 1070, each with 8GiB memory GPU driver: 418.87.01, CUDA: 10.1 It took about 3.5 hours to finish. Note that the above results didn't involve any hyperparameter search.
MobileBERT is a thin version of BERTLARGE, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks. This model was fine-tuned from the HuggingFace checkpoint google/mobilebert-uncased on SQuAD1.1. CPU: Intel(R) Core(TM) i7-6800K CPU @ 3.40GHz Memory: 32 GiB GPUs: 2 GeForce GTX 1070, each with 8GiB memory GPU driver: 418.87.01, CUDA: 10.1 It took about 3 hours to finish. Note that the above results didn't involve any hyperparameter search.
