A method, device and equipment for constructing a military field knowledge extraction large model
By optimizing the large language model through a multi-stage training framework and a composite reward function, the problem of low training efficiency in military knowledge extraction tasks is solved, and the ability to extract knowledge from basic to advanced levels is improved, especially in the learning of high-value, low-frequency knowledge.
CN121882037BActive Publication Date: 2026-06-02XIAMEN YUANTING INFORMATION TECH CO LTD
Patent Information
- Authority / Receiving Office
- CN Β· China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIAMEN YUANTING INFORMATION TECH CO LTD
- Filing Date
- 2026-03-19
- Publication Date
- 2026-06-02
Smart Images

Figure CN121882037B_ABST
Abstract
The application discloses a kind of military field knowledge extraction large model construction method, device and equipment, comprising: obtaining military field original text and pre-processing and marking, generate training sample;According to corresponding target complexity interval from training sample corresponding sample is screened, constitute the candidate training set of current training stage;Based on dynamic data matching strategy from candidate training set is sampled to construct training batch, to large language model is supervised fine-tuning;Knowledge extraction is carried out to input text using the model after supervision fine-tuning, obtain knowledge extraction result, calculate knowledge extraction result using composite reward function, obtain reward value;Based on reward value, the model parameter of model is reinforced learning optimization, and the performance of model is periodically evaluated;After judging the performance of model in current training stage reaches convergence condition, the model parameter of model is used as the initial parameter of next training stage and continues to iterate training, obtains military field knowledge extraction large model.
Need to check novelty before this filing date? Find Prior Art
Citation Information
Patent Citations
CN117874252A
CN121434415A