Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support, with the following key features: - Uniquely support of seamless switching between thinking mode (for complex logical reasoning, math, and coding) and non-thinking mode (for efficient, general-purpose dialogue) within single model, ensuring optimal performance across various scenarios. - Significantly enhancement in its reasoning capabilities, surpassing previous QwQ (in thinking mode) and…
NER_Game is an open-weight model for text generation from Hongrui, released under Apache License 2.0. It has 4B parameters and a 40,960-token context. At 16-bit it needs about 9.7 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.
游戏 UGC 评论「语义单元抽取」模型。基于 Qwen/Qwen3-4B，LoRA（r16 / alpha32 / dropout0.05）训练后合并进底座，fp16 权重，可直接用 transformers 加载，无 peft 依赖。 - 任务：从一条评论文本中抽取语义单元（mention / span / occurrence），只做抽取，不做分类 - 硬件：实测峰值显存 8.82 GB，T4/T10（16GB，Turing）可直接部署，attention 走 sdpa，无需 FlashAttention2…
Parameters4B
Context40,960
Weights16.1 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—
Runs On
What it takes to serve NER_Game (4B 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 | 8.0 GB | 9.7 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 4.0 GB | 4.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 2.0 GB | 2.4 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 18, 2026.
NER_Game on every accelerator the SAVRN Index prices, at every precision
Model Card
By Hongrui, published under apache-2.0, revision 43bc56ac7e48.
游戏 UGC 评论「语义单元抽取」模型。基于 Qwen/Qwen3-4B，LoRA（r16 / alpha32 / dropout0.05）训练后合并进底座，fp16 权重，可直接用 transformers 加载，无 peft 依赖。 - 任务：从一条评论文本中抽取语义单元（mention / span / occurrence），只做抽取，不做分类 - 硬件：实测峰值显存 8.82 GB，T4/T10（16GB，Turing）可直接部署，attention 走 sdpa，无需 FlashAttention2 生成参数已固化在 generationconfig.json：贪心解码（dosample=False）+ maxnewtokens=2048，请勿改动，否则与评测结果不一致。 ok=False 表示输出 JSON 解析失败或存在非法 span（mention 不在正文 / occurrence 越界），此时 semanticunits 为已通过校验的子集；线上建议对 ok=False 记录走兜底（重试或降级）。 - JSON 成功率 98.99%，非法 span 0 - SDK 与评测输出平价验证：20/20 DEV 样本单元集合完全一致 1. dtype 必须 fp16：Turing 架构无 bf16 支持；Ampere+（A10/A100/4090）可改用 bf16 3. 时延：A100 mean 4.8s / p50 3.8s 每条；Turing 预计 8~15s，随单元数线性增长 4. 输入 text 建议 ≤ 1024 token（训练 maxseq…
Read Hongrui's full model card
extraction-r19-v1
游戏 UGC 评论「语义单元抽取」模型。基于 Qwen/Qwen3-4B，LoRA（r16 / alpha32 / dropout0.05）训练后合并进底座，fp16 权重，可直接用 transformers 加载，无 peft 依赖。
- 任务：从一条评论文本中抽取语义单元（mention / span / occurrence），只做抽取，不做分类
- 权重：fp16，4.02B 参数，约 7.5 GB
- 硬件：实测峰值显存 8.82 GB，T4/T10（16GB，Turing）可直接部署，attention 走
sdpa，无需 FlashAttention2
目录结构
.
├── model/ # 合并后的完整模型（transformers 可直接加载）
│ ├── config.json
│ ├── model-0000{1,2}-of-00002.safetensors
│ ├── model.safetensors.index.json
│ ├── generation_config.json
│ ├── chat_template.jinja # 提示词契约已固化在模板中
│ └── tokenizer 相关文件
├── sdk/extraction_sdk/ # 自包含 SDK（契约 + 抽取器 + 后处理校验）
├── example.py
├── manifest.json # 指标 / 契约 / 硬件要求
├── SHA256SUMS.txt # 权重文件校验值
└── RELEASE.md
快速开始
1. 用 SDK（推荐，含输出校验与契约）
pip install torch>=2.1 transformers>=4.53
git clone https://huggingface.co/<YOUR_NAMESPACE>/<YOUR_REPO> extraction-r19-v1
cd extraction-r19-v1
python example.py
import sys; sys.path.insert(0, "sdk")
from extraction_sdk import SemanticUnitExtractor
ext = SemanticUnitExtractor("model", dtype="fp16") # T4/T10 必须 fp16；Ampere+ 可用 bf16
out = ext.extract("策划在偷偷加强小兵,我情绪稳定#王者荣耀",
record_id="r1", game_name="王者荣耀")
# out = {"id","text","ok","latency_ms","semantic_units":[{"mention","start","end","unit_id",...}],"diagnostics":{...}}
2. 直接用 transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("<YOUR_NAMESPACE>/<YOUR_REPO>", subfolder="model")
model = AutoModelForCausalLM.from_pretrained(
"<YOUR_NAMESPACE>/<YOUR_REPO>", subfolder="model",
dtype="fp16", attn_implementation="sdpa", device_map="auto",
)
# 提示词请走 tokenizer 内置的 chat template（提示词契约 reasoningner-lite-sft-v5-extraction-only）
生成参数已固化在 generation_config.json：贪心解码（do_sample=False）+ max_new_tokens=2048，请勿改动，否则与评测结果不一致。
输入 / 输出契约
| 契约 | 版本 |
|---|---|
| 提示词 | reasoningner-lite-sft-v5-extraction-only |
| 模型输出 | semantic-units-occurrence-v3-extraction-only |
| 正式预测 | semantic-units-span-v3-extraction-only |
ok=False 表示输出 JSON 解析失败或存在非法 span（mention 不在正文 / occurrence 越界），此时 semantic_units 为已通过校验的子集；线上建议对 ok=False 记录走兜底（重试或降级）。
评测（DEV 298 条，人工裁决金标）
| 口径 | P | R | F1 | Δ vs 旧生产模型 R1 |
|---|---|---|---|---|
| 严格 mention（门控主口径） | 0.7376 | 0.7855 | 0.7608 | +0.0174（95% CI [0.0013, 0.0336]） |
| occurrence 逐位置 | 0.7273 | 0.7752 | 0.7505 | — |
| relaxed（后缀 + 有界包含） | 0.7915 | 0.8429 | 0.8164 | +0.0169 |
- JSON 成功率 98.99%，非法 span 0
- 泛化性（剔除政策词表）Δ +0.0133
- SDK 与评测输出平价验证：20/20 DEV 样本单元集合完全一致
部署要点
- dtype 必须 fp16：Turing 架构无 bf16 支持；Ampere+（A10/A100/4090）可改用
bf16 - 显存峰值 8.82 GB，16GB 卡可并发 1~2 实例
- 时延：A100 mean 4.8s / p50 3.8s 每条；Turing 预计 8~15s，随单元数线性增长
- 输入
text建议 ≤ 1024 token（训练 max_seq 993） - 分类（
semantic_type等四维）不在本包内，属后置阶段
已知限制
- 2/298 条顽固失败模式：模型幻觉 mention 不在正文（走
ok=False兜底） - 基于 Qwen3-4B，继承其偏见与事实性风险，输出需结合业务审核
- 领域为游戏 UGC 中文评论，跨领域效果未验证
许可证
Apache-2.0，沿用底座模型 Qwen/Qwen3-4B 的许可证。
训练信息
- 底座：Qwen/Qwen3-4B
- 微调：LoRA r16 / alpha32 / dropout0.05，合并进底座
- 训练数据：2384 条 / 8863 单元（政策对齐 v3）
- 精度：float16
Configuration
- Architecture
- Qwen3ForCausalLM
- Context length (tokens)
- 40,960
- Layers
- 36
- Hidden size
- 2,560
- Feed-forward size
- 9,728
- Attention heads
- 32
- Key/value heads
- 8
- Head dimension
- 128
- Vocabulary size
- 151,936
- RoPE base
- 1,000,000
- Stored precision
- float16
- Model type
- qwen3
Identity and Version
- Repository
- li-ovo/NER_Game
- Publisher
- Hongrui
- Task
- Text generation
- Modality
- Text
- Library
- transformers
- Parameters
- 4B parameters
- Languages
- zh
- Revision
- 43bc56ac7e48b2cf79bf555525742cf0467d1bc8
- First published
- 2026-09-22
- Last updated
- 2026-09-22
Files and Weights
34 files, 16.1 GB in total. The weights are 4 files totalling 16.1 GB in safetensors.
Weights4 files · 16.1 GB
Configuration15 files · 84.7 KB
Tokenizer8 files · 31.8 MB
Documentation2 files · 7.5 KB
Other4 files · 8.8 KB
Repository1 file · 1.6 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00002.safetensors | Weights | 5.0 GB | 1ad2947f50f7 |
| model-00002-of-00002.safetensors | Weights | 3.1 GB | 174097609d19 |
| model/model-00001-of-00002.safetensors | Weights | 5.0 GB | 1ad2947f50f7 |
| model/model-00002-of-00002.safetensors | Weights | 3.1 GB | 174097609d19 |
| added_tokens.json | Configuration | 707 B | — |
| config.json | Configuration | 1.5 KB | — |
| example.py | Configuration | 835 B | — |
| generation_config.json | Configuration | 214 B | — |
| manifest.json | Configuration | 1.9 KB | — |
| model.safetensors.index.json | Configuration | 32.9 KB | — |
| model/added_tokens.json | Configuration | 707 B | — |
| model/config.json | Configuration | 1.5 KB | — |
| model/generation_config.json | Configuration | 214 B | — |
| model/model.safetensors.index.json | Configuration | 32.9 KB | — |
| model/special_tokens_map.json | Configuration | 613 B | — |
| sdk/extraction_sdk/__init__.py | Configuration | 321 B | — |
| sdk/extraction_sdk/contracts.py | Configuration | 1.8 KB | — |
| sdk/extraction_sdk/extractor.py | Configuration | 8.0 KB | — |
| special_tokens_map.json | Configuration | 613 B | — |
| README.md | Documentation | 4.7 KB | — |
| RELEASE.md | Documentation | 2.9 KB | — |
| SHA256SUMS.txt | Other | 357 B | — |
| chat_template.jinja | Other | 4.2 KB | — |
| model/chat_template.jinja | Other | 4.2 KB | — |
| requirements.txt | Other | 62 B | — |
| .gitattributes | Repository | 1.6 KB | — |
| merges.txt | Tokenizer | 1.7 MB | — |
| model/merges.txt | Tokenizer | 1.7 MB | — |
| model/tokenizer.json | Tokenizer | 11.4 MB | aeb13307a71a |
| model/tokenizer_config.json | Tokenizer | 5.4 KB | — |
| model/vocab.json | Tokenizer | 2.8 MB | — |
| tokenizer.json | Tokenizer | 11.4 MB | aeb13307a71a |
| tokenizer_config.json | Tokenizer | 5.4 KB | — |
| vocab.json | Tokenizer | 2.8 MB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 16.1 GB
Download from Hongrui
Released by Hongrui through its official repository on Hugging Face. Read the license.
Built From
- Adapter of Qwen/Qwen3-4B
- Derived from Qwen/Qwen3-4B
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 16.1 GB |
| 16-bit | 8.0 GB |
| 8-bit | 4.0 GB |
| 4-bit | 2.0 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About NER_Game
How much GPU memory does NER_Game need?
About 9.7 GB at 16-bit and 2.4 GB at 4-bit: the weights (4B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run NER_Game 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 NER_Game commercially?
Yes. NER_Game is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
What is NER_Game's context length?
40,960 tokens, from the maximum position embeddings in its published configuration.
Similar Models
We introduce the updated version of the Qwen3-4B non-thinking mode, named Qwen3-4B-Instruct-2507, featuring the following key enhancements: - Significant improvements in general capabilities, including instruction following, logical reasoning, text comprehension, mathematics, science, coding and tool usage. - Substantial gains in long-tail knowledge coverage across multiple languages. - Markedly better alignment with user preferences in subjective and open-ended tasks, enabling more helpful responses and higher-quality text generation. - Enhanced capabilities in 256K long-context understanding. Qwen3-4B-Instruct-2507 has the following features: NOTE: This model supports only non-thinking…
Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Building upon extensive advancements in training data, model architecture, and optimization techniques, Qwen3 delivers the following key improvements over the previously released Qwen2.5: Qwen3-4B-Base has the following features: For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our blog, GitHub, and Documentation. The code of Qwen3 has been in the latest Hugging Face transformers and we advise you to use the latest version of transformers. With transformers<4.51.0, you will…
Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Building upon extensive advancements in training data, model architecture, and optimization techniques, Qwen3 delivers the following key improvements over the previously released Qwen2.5: Qwen3-4B-Base has the following features: For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our blog, GitHub, and Documentation. The code of Qwen3 has been in the latest Hugging Face transformers and we advise you to use the latest version of transformers. With transformers<4.51.0, you will…
SRA-RiskGate-4B is an autonomous risk scoring, compliance verification, and dispute adjudication model fine-tuned on top of Qwen/Qwen3-4B-Instruct-2507. It is engineered for both ends of a stablecoin payment's operational lifecycle: Autonomous on-chain agents can hallucinate payment transfers, sign malformed calldata, or trigger catastrophic transactions during market depegs. The deterministic pre-filter and policy firewall is available directly as a verified Action Provider for the Coinbase AgentKit framework: Register RiskGateActionProvider as the payment provider on your AgentKit instance. It checks chain support, policies, and peg deviations, executing safe ERC-20 transfers only after…
![[HYU_NLP] EVA Team logo](/media/mh-logo-hyu-nlp-eval-558e075b88.webp){kind=logo}
Intermediate policy from dynamic OnlineRubrics-Every GRPO training. Distinct from static-rubric GRPO. Base model: Qwen/Qwen3-4B-Instruct-2507; thinking disabled. This checkpoint is a policy state used by the Phase-1 audit. No downstream medical capability or safety claim is made. Research use only; not validated for clinical decision-making. Root files are the veRL-exported Hugging Face inference model (BF16). originalcheckpoint/ preserves the exact original FSDP parameter checkpoint and tokenizer/configuration files. Optimizer state, training data, responses, rubrics, infrastructure configuration, and credentials are not included. The original is retained because export…