Thai extractive QA fine-tuned from airesearch/wangchanberta-base-att-spm-uncased with SimpleTransformers (10301374 NLP final exam, Q07).
SAVRN Model Hub · Models by Task
Question Answering Models
73 open-weight question answering models in the SAVRN Model Hub, with Deepset, Team Mradermacher and Qingqing Cao publishing the most.
73Models
43Publishers
14M to 1.2BParameter range
6Licenses
SAVRN's Take
Most of what our catalog files under question answering is extractive work: give the model a passage and a question, and it marks the span that answers it, or says nothing there does. Deepset's SQuAD2.0 fine-tunes, trained with unanswerable questions included, set the pattern, and 17 of the 53 entries carry that name. Sizes run from 14 million parameters for koelectra-small-v2-distilled-korquad-384 to 560 million for xlm-roberta-large-squad2, and every one reads 512 or 514 tokens at a time. That window is the real constraint: the model sees one passage per call, so the retrieval in front of it decides more than the weights do.
Memory is nearly a non-issue. The 560M model needs 1.3 GB at 16-bit and 0.3 GB at 4-bit. roberta-base-squad2, the most downloaded entry with live pricing at 559,658 pulls a month, needs 0.3 GB at 16-bit. Both list the same cheapest host, one MI300X at $1.85 an hour, so the buying question becomes how many copies you stack on a card you already own.
Licensing is where we slow down. MIT covers 15 entries and Apache-2.0 another 11, both permissive. CC-BY-4.0 covers 12, the Deepset models leading the downloads among them, which means crediting the publisher wherever the model runs, and two entries carry the Llama 3.1 license. And 13 of the 53 state no license, including the most downloaded model, DNeff's electra_large_discriminator_squad2_512 at 768,739 downloads a month. Confirm the license file with the publisher before you commit, and reach for tinyroberta-squad2 or the Apache-2.0 distilbert-base-cased-distilled-squad when the goal is many instances on one card.
Most Downloaded
| Model | Publisher | Parameters | License | Monthly downloads | Cheapest GPUs at 16-bit |
|---|---|---|---|---|---|
| electra_large_discriminator_squad2_512 | DNeff | — | Not stated | 768.7k | — |
| roberta-base-squad2 | Deepset | 124M | cc-by-4.0 | 559.7k | 1x MI300X, $1.85/hr |
| deberta-v3-base-squad2 | Deepset | 184M | cc-by-4.0 | 247.6k | 1x MI300X, $1.85/hr |
| bert-large-uncased-whole-word-masking-squad2 | Deepset | 335M | cc-by-4.0 | 206.1k | 1x MI300X, $1.85/hr |
| distilbert-base-cased-distilled-squad | DistilBERT community | 65M | apache-2.0 | 198.9k | 1x MI300X, $1.85/hr |
| tinyroberta-squad2 | Deepset | 82M | cc-by-4.0 | 183.5k | 1x MI300X, $1.85/hr |
| koelectra-small-v2-distilled-korquad-384 | Jangwon Park | 14M | Not stated | 163.1k | 1x MI300X, $1.85/hr |
| splinter-base | Tel Aviv University | — | apache-2.0 | 104.9k | — |
| koelectra-base-v3-finetuned-korquad | Jangwon Park | 112M | Not stated | 103k | 1x MI300X, $1.85/hr |
| bert-base-cased-squad2 | Deepset | 108M | cc-by-4.0 | 71.7k | 1x MI300X, $1.85/hr |
Licenses
| License | Models | Commercial use |
|---|---|---|
| not stated | 22 | Not stated |
| mit | 16 | Yes |
| apache-2.0 | 16 | Yes |
| cc-by-4.0 | 13 | Yes |
| other | 4 | Read the license |
| llama3.1 | 2 | Yes, with conditions |
Who Publishes Them
| Publisher | Models |
|---|---|
| Deepset | 17 |
| Team Mradermacher | 6 |
| Qingqing Cao | 3 |
| VietAI | 2 |
| Benjamin Horn | 2 |
| Manuel Romero | 2 |
All 73 Models, Page 2 of 2
This is the model card of a transformers model that has been pushed on the Hub. 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).
This is the model card of a transformers model that has been pushed on the Hub. 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).
This is the model card of a transformers model that has been pushed on the Hub. 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).
Thai extractive QA fine-tuned จาก airesearch/wangchanberta-base-att-spm-uncased ด้วย 4 ตัวอย่าง (engine: transformers, steps: 30) ข้อสอบปลายภาควิชา 10301374 โดย 6704101346 ข้อมูลฝึกมีเพียง 4 คู่คำถาม-คำตอบ โมเดลจึงจดจำรูปแบบเฉพาะ (overfit) มากกว่าเรียนรู้การอ่านเข้าใจ ผลบนข้อมูลที่ฝึกจึงไม่สะท้อนความสามารถจริง ไม่มีชุด validation ที่เป็นอิสระ ค่าที่วัดได้แปรปรวนสูงตาม seed และคำถามส่วนใหญ่เป็นแบบ 'อยู่ภาคใด' ทำให้โมเดลอาจเดาจากรูปแบบ ควรใช้ชุดข้อมูลขนาดใหญ่ เช่น thaiqasquad หรือ iappwikiqasquad และแยก train/validation/test ก่อนสรุปผล
This is the model card of a transformers model that has been pushed on the Hub. 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).
This is the model card of a transformers model that has been pushed on the Hub. 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).
[2026-10-06] Our paper is available on arXiv. - [2026-10-06] Code released on GitHub, and model released on HuggingFace. - [2026-10-06] Project website is live! OpenSLA is a new family of models - Sensor-Langauge-Action (SLA) models. It takes a multi-channel sensor history together with language context and produces a structured action prediction and a sensor-state caption. We release prompt as a flat token prefix. the signal into local, per-channel, and global memory tokens. - Python >= 3.10 - PyTorch >= 2.5 - Transformers >= 5.8.1 (for the Qwen3.5 backbone) - PEFT >= 0.12 demo.ipynb loads a checkpoint, builds an input batch, and runs prediction. Download the files from this repository and…
Oracle7 is an AI built to say "I can't verify that" rather than guess. It is not a neural language model. It answers from stored, sourced records and from 267 specialists, each of which passed its own held-out test before it was allowed to answer. Every reply carries its path: the route, specialists and records that produced it. When nothing in it can verify an answer, it declines and tells you why. It does not always manage that: the measured section below says exactly how often it is still wrong, and on what. This is the full release: the core plus the deep knowledge stores (English Wikipedia + DBpedia abstracts, 71 OpenStax textbooks, Simple English Wikipedia, Wiktionary, WordNet…
Oracle7 is an AI built to say "I can't verify that" rather than guess. It is not a neural language model. It answers from stored, sourced records and from 267 specialists, each of which passed its own held-out test before it was allowed to answer. Every reply carries its path: the route, specialists and records that produced it. When nothing in it can verify an answer, it declines and tells you why. It does not always manage that: the measured section below says exactly how often it is still wrong, and on what. This is the slim release: the same core, routes and 267 specialists, with the smaller stores (Simple English Wikipedia, Wiktionary, WordNet, ConceptNet, Wikidata relations) and no…
Second pre-alpha from Tyness Train, an agent-guided fine-tuning harness for small LLMs on low-VRAM GPUs. This model is RAG-style grounded: every prompt carries excerpts of the relevant norm, and the model is trained to answer only from that context. Three iterations established that SFT alone teaches format, not facts on 1.2B models: legalize-es, with teacher-generated ChatML entries from es-legal-priv. 10 BOE articles, 1,148 train + 20 held-out eval items: - Each entry cites specific precepts (articuloscitados); those sections are extracted from the norm markdown and embedded as a CONTEXTO block in the user turn (up to 9,000 chars). - The system prompt enforces grounded behavior: answer…
Third pre-alpha from Tyness Train, an agent-guided fine-tuning harness for small LLMs on low-VRAM GPUs. This model is RAG-style grounded (answers only from provided norm excerpts) and adds two confirmed recipe upgrades: a structured METADATOS block in the system prompt (official identifiers, dates, rank, department, article-title index) and adversarial correction pairs (user asserts a false premise → the model corrects it with the exact citation). Evaluated on fresh, uncontaminated teacher-generated sets: 200 questions per context style (never part of any training split), each grounded in a specific precept, judged by an LLM referee (glm5.3-flash): correct / partial / incorrect + 0–1 score.…
This is the model card of a transformers model that has been pushed on the Hub. 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).
Questions
Which Question answering models are most downloaded?
By monthly downloads reported by the Hugging Face Hub: .