AgriJev: Efficient Agricultural Candidate Scoring with Multi-Scale Adapted Models

Xuechen Li

cs.CLOct 14, 2026

AgriJev: Efficient Agricultural Candidate Scoring with Multi-Scale Adapted Models

Authors: Xuechen Li

Organizations: University of Minnesota Twin Cities

Abstract

Agricultural systems make decisions over heterogeneous entity types, textual judgments, and knowledge-question options. Generating a textual answer adds decoding and extraction steps to decisions that ultimately require a candidate choice. We introduce AgriJev, a family of adapted Qwen3.5 models at 0.8B, 2B, 4B, and 9B. A shared interface conditions on agricultural text, a judgment question, and candidate definitions, returning scores and relative probabilities in one forward call. We prepare source-aware training records and frozen evaluations from existing agricultural resources, retaining original groups and annotation provenance. Comparisons include original Qwen3.5 models and matched agricultural SFT models, both evaluated through native answer generation. Across 17 external test views, AgriJev-9B reaches 75.44% mean accuracy, exceeding the original and SFT generators by 5.89 and 3.79 percentage points. It outperforms the original generator on 13 views and narrows the mean gap to hosted Jev 1.13.0. On a fixed A100 workload, its median scoring latency is approximately half that of same-scale answer generation. Task-dependent gains and failures characterize a practical choice among agricultural model weights and inference contracts.

Figures & tables

Figure 1: Accuracy, latency, and model size. Mean accuracy uses 17 external views, excluding paired CROP reordering. p50 latency uses 32 BIO texts, 27 candidates, A100 40GB, batch 1. Qwen3.5 and Qwen3.5-SFT generate answers; AgriJev scores candidates. Bubble area is proportional to measured loaded parameters; the lower-right size key gives nominal scales. Accuracy and latency use distinct fixed workloads.

Figure 2: AgriJev candidate scoring and adaptation. Agricultural text, a question, and candidate definitions form a prompt with locally assigned option codes. Trainable query/value LoRA augments frozen Qwen3.5 weights; the original LM head supplies vocabulary logits in one forward pass. Scores for the current candidate codes are gathered and normalized within that set. The highest-scoring code identifies the selected option (B, corresponding to Label 2, in this illustration). The dashed red path denotes candidate-set cross-entropy and backpropagation to LoRA during training. Correct answers provide supervision and are excluded from the input prompt.

Family	Model	Typing	Assertion	Event	BIO accuracy	BIO macro-F1
AgriJev (ours)	0.8B	96.17	89.83	80.25	92.75	79.63
	2B	96.70	95.46	84.00	94.03	83.77
	4B	97.12	95.89	84.15	94.46	84.33
	9B	97.62	95.30	81.93	95.57	86.37
Qwen3.5 ( 2026 )	0.8B	96.17	87.21	73.35	93.01	79.86
	2B	97.07	95.28	83.46	94.63	82.75

Table 1: Supervised internal evaluation (%). All models use the 41,724-record training release; Qwen3.5 rows use SFT and native answer generation, while AgriJev and encoder controls score candidates. Model sizes are nominal. Family citations are given in Section 4.2 . Typing, assertion, and event views number 3,578, 7,156, and 1,306; events are silver. BIO is a 1,173-record internal subset with fixed 27-class macro-F1. Bold marks the highest accuracy or macro-F1 in each column, including ties.

Dataset	Jev ( 2026 )	AgriJev (ours)	Qwen3.5 ( 2026 )	Qwen3.5-SFT ( 2026 )	AgriReason ( 2026 )	GLiClass ( 2025 )	NLI ( 2023 )	BGE ( 2024 )	Akshara ( 2024 )
AgriEval	67.60	62.20	58.18	62.30	37.29	26.21	40.23	43.02	42.49
CROP	93.52	91.10	90.18	93.02	45.19	19.31	42.38	28.40	66.53
CMMLU agronomy	87.57	76.33	72.19	76.92	28.40	26.63	23.67	30.77	38.46
TMMLU+ agriculture	76.39	65.97	52.78	61.81	23.61	29.86	22.22	22.22	27.78
C-Eval plant protection	90.45	83.42	80.40	82.91	43.22	30.15	27.14	25.13	38.69
C-Eval veterinary	92.38	91.43	83.81	90.48	40.48	30.48	27.14	29.52	34.76

Table 2: Zero-shot external accuracy (%). Qwen3.5, Qwen3.5-SFT, and AgriJev use 9B; AgriReason uses 4B, Akshara 7B, and Jev version 1.13.0. Both Qwen conditions generate answers; AgriJev scores candidates. Header citations identify model families; SFT follows Section 4.2 . Dataset citations are in Section 4.1 . SeedBench averages five view accuracies equally;

\dagger

marks BGE averages using qualified subsets for three reading views, detailed in Appendix D . Bold marks only the highest complete-coverage result in each row, including ties. Paired CROP reordering and full view results appear in Appendix F .

Figure 3: Observed deployment latency. (a) Points give p50 and right-hand caps p95 on 32 BIO texts with 27 candidates, A100 40GB, batch 1, and 96 timings per model. Both Qwen families generate answers; AgriJev scores candidates. (b) Hosted Jev ranges summarize seven external views and include service, network, waiting, and retries. Panels have different inputs and timing conditions.

Figure 4: External capability profiles. Seven source-defined blocks contain 4, 1, 3, 5, 1, 2, and 1 views. Each axis averages complete view accuracies on the same 0–100% scale. Qwen3.5 and Qwen3.5-SFT use native generation; AgriJev uses candidate scoring; Jev uses hosted choices. Block definitions appear below and in Appendix H . Paired CROP reordering is excluded.

Figure 5: Actual decisions on two external questions. Original Chinese inputs are translated only for display. Rows retain saved AgriJev, original-Qwen generation, SFT generation, and Jev predictions on identical IDs. Outlined codes match the source key. No confidence is inferred for generated answers. The original two selected questions illustrate successes and failures; they do not estimate their frequency.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Prepared-release composition. (a) The shared 41,724-record training release combines 10,654 AgriCHN source-BIO examples, 18,828 CROP-generated entity examples, and 12,242 native BBK English choices. (b) The 12,040 internal views contain typing, assertion, and event tasks. Counts are exact frozen-release totals, not independent annotation counts: typing and assertions share source labels, and BIO is a typing subset.

Source	Records	Groups	Supervision
AgriCHN	10,654	2,135	Source BIO
CROP entities	18,828	4,853	Generated labels
BBK English	12,242	12,111	Native answers
Total	41,724	Two training task families

Appendix

Table 3: Prepared training records and preserved source-group IDs. Group counts identify provenance units within each source, not newly independent annotations. CROP entity labels remain generated supervision.

Stage	Rows
Released English rows	12,648
Duplicate-option exclusion	259
Empty-option exclusion	3
Duplicate-question exclusion	143
External exact-overlap exclusion	1
Accepted training records	12,242

Appendix

Table 4: BBK English filtering, applied before adding native choices to the training release. Each exclusion count is recorded by the frozen preparation audit.

Task view	Records	Groups
Entity typing	3,578	877
Type assertion	7,156	877
Event presence	1,306	638
All internal views	12,040	1,505
BIO subset	1,173	251

Appendix

Table 5: Internal task-view inventory. BIO is a subset of typing. Typing and assertion group counts overlap; counts across these rows must not be added to infer independent groups.

Setting	Decoders	Encoders
Training records	41,724	41,724
Development records	600	600
Epochs / updates	1 / 2,608	1 / 2,608
Random seed	13	13
Effective batch	16	16
Learning rate	$10^{-4}$	$2\times 10^{-5}$

Appendix

Table 6: Frozen training settings. AgriJev and native-answer SFT share this decoder schedule; their supervised outputs differ. Encoder controls share data and update count but use a different optimizer learning rate and checkpoint-selection schedule.

Scale	Minutes	Dev. response CE
0.8B	64.42	0.2160
2B	37.59	0.1178
4B	89.45	0.1074
9B	49.64	0.1246

Appendix

Table 7: Completed native-answer SFT runs. Minutes denote recorded training time; development CE is teacher-forced response loss at the fixed-epoch checkpoint. These timings are not inference latency.

Scale	Loaded	Trainable
0.8B	853,624,896	638,976
2B	2,214,077,248	835,584
4B	4,541,100,544	1,835,008
9B	9,412,041,968	2,228,224

Appendix

Table 8: Actual loaded and trainable parameter counts for each decoder scale. AgriJev and native-answer SFT at the same scale share these counts; nominal size labels are not substituted for the measured totals.

Model	Mode	Accuracy	Macro-F1
Qwen3.5-0.8B	Generate	31.12	22.67
Qwen3.5-2B	Generate	52.00	39.25
Qwen3.5-4B	Generate	80.82	70.35
Qwen3.5-9B	Generate	86.45	78.43
Qwen3.5-0.8B	Score	38.02	29.58
Qwen3.5-2B	Score	66.07	57.31

Appendix

Table 9: Off-the-shelf BIO references (%). Generate denotes native answer generation; Score retains the prior candidate-scoring contract. Identical source records and the fixed 27-class metric are used; the two contracts are not interchangeable.

Original class (English gloss)	Support	Qwen gen.	SFT gen.	AgriJev	Token-SFT
Time	52	93.33	94.34	95.24	95.24
Soil moisture	9	94.74	94.74	100.00	100.00
Aquatic products	4	100.00	100.00	100.00	100.00
Fruits	40	89.89	100.00	100.00	96.20
Diseases, pests, weeds	203	85.71	98.26	98.52	98.52
Vegetables	59	84.38	92.19	89.39	90.08

Appendix

Table 10: BIO class-level F1 (%) at 9B. Qwen and native SFT generate answers; AgriJev and prior token-SFT score candidates. Support counts gold records. English names are display glosses, not relabeled classes.

Scale	Generator	AgriEval	Seed 1-2	Seed 3-2
0.8B	Original	28.75	3.44	4.95
0.8B	Native SFT	29.22	7.22	8.91
2B	Original	28.85	4.12	7.92
2B	Native SFT	35.78	9.62	10.40
4B	Original	6.04	3.78	7.92
4B	Native SFT	44.69	5.50	14.36

Appendix

Table 11: Reconstructed multi-select exact match (%) from native generated option decisions: 1,920 AgriEval, 291 SeedBench 1-2, and 202 SeedBench 3-2 original questions. All options must be correct. These are distinct from decision accuracies.

Model	Inference contract	p50	p95	Throughput
Qwen3.5-0.8B	Original generation	124.04	125.51	8.06
Qwen3.5-0.8B-SFT	SFT generation	127.58	131.53	7.82
AgriJev-0.8B	Candidate scoring	42.96	43.62	23.25
Qwen3.5-2B	Original generation	125.17	128.56	7.97
Qwen3.5-2B-SFT	SFT generation	131.56	132.42	7.60
AgriJev-2B	Candidate scoring	41.90	70.84	21.95

Appendix

Table 12: Current local end-to-end timings under the declared native contracts. Quantiles are milliseconds; throughput is records/s. Hardware and source inputs are fixed, but prompts and returned objects differ. All quantiles derive from 96 recorded request durations.

Dataset	Records	Jev 1.13.0	AgriJev (ours)	Qwen3.5 original	Qwen3.5 SFT	AgriReason 4B	GLiClass	NLI	BGE	Akshara 7B
AgriEval	20,738	67.60	62.20	58.18	62.30	37.29	26.21	40.23	43.02	42.49
CROP	5,043	93.52	91.10	90.18	93.02	45.19	19.31	42.38	28.40	66.53
CROP reordered	5,043	–	91.71	91.31	93.16	46.04	21.08	42.59	28.36	69.05
CMMLU agronomy	169	87.57	76.33	72.19	76.92	28.40	26.63	23.67	30.77	38.46
TMMLU+ agriculture	144	76.39	65.97	52.78	61.81	23.61	29.86	22.22	22.22	27.78
C-Eval plant protection	199	90.45	83.42	80.40	82.91	43.22	30.15	27.14	25.13	38.69

Appendix

Table 13: Complete external accuracy (%). Qwen3.5, native-answer Qwen3.5-SFT, and AgriJev use 9B; the other nominal scales are shown in the header. CROP reordered is paired and excluded from the mean. Bold marks only the highest complete-coverage result per row, including ties. BGE

\dagger

entries use 227/264, 612/808, and 214/279 qualified records in SeedBench 3-1, 3-2, and 3-5. Dashes denote unavailable full-view results. Data and model sources are cited in Sections 4.1 and 4.2 .

Dataset	0.8B Qwen gen.	0.8B SFT gen.	0.8B AgriJev	2B Qwen gen.	2B SFT gen.	2B AgriJev
AgriEval	46.62	49.00	48.79	47.05	51.08	50.60
CROP	80.88	84.24	84.31	85.48	87.82	86.62
CROP reordered	81.30	84.71	83.66	86.24	88.74	87.15
CMMLU agronomy	37.87	43.20	46.15	47.34	53.85	53.85
TMMLU+ agriculture	20.14	30.56	24.31	29.86	34.03	38.19
C-Eval plant protection	44.22	50.75	48.74	56.78	63.32	63.32

Appendix

Table 14: Complete 0.8B/2B external accuracy (%). Qwen and native-answer SFT generate responses; AgriJev scores candidates. All adapted conditions use the current 41,724-record release. CROP reordered is paired and omitted from the 17-view mean.

Dataset	4B Qwen gen.	4B SFT gen.	4B AgriJev	9B Qwen gen.	9B SFT gen.	9B AgriJev
AgriEval	44.00	57.72	58.51	58.18	62.30	62.20
CROP	89.75	91.65	91.75	90.18	93.02	91.10
CROP reordered	90.40	91.71	91.87	91.31	93.16	91.71
CMMLU agronomy	61.54	68.05	68.64	72.19	76.92	76.33
TMMLU+ agriculture	47.92	50.69	51.39	52.78	61.81	65.97
C-Eval plant protection	75.38	77.89	80.40	80.40	82.91	83.42

Appendix

Table 15: Complete 4B/9B external accuracy (%). Qwen and native-answer SFT generate responses; AgriJev scores candidates. All adapted conditions use the current 41,724-record release. CROP reordered is paired and omitted from the 17-view mean.

Dataset	GLiClass-FT	RoBERTa-FT
AgriEval	44.61	39.10
CROP	70.61	36.17
CMMLU agronomy	30.77	25.44
TMMLU+ agriculture	29.17	28.47
C-Eval plant protection	31.66	26.63
C-Eval veterinary	30.95	30.00

Appendix

Table 16: External transfer accuracy (%) for trained encoder controls. All rows have complete frozen-set coverage. The final mean excludes paired CROP reordered; these columns are separate from the off-the-shelf GLiClass column in Table 13 .

Scale	Generator	Invalid	Record rate (%)	Largest-fraction view	Invalid/records	View-mean rate (%)	Repair bound mean (%)
0.8B	Original	98	0.31	SeedBench 3-2	82/808	0.69	48.10
0.8B	Native SFT	155	0.49	SeedBench 3-2	122/808	1.03	54.65
2B	Original	226	0.71	SeedBench 3-2	178/808	1.54	54.29
2B	Native SFT	332	1.04	SeedBench 3-2	230/808	2.18	61.10
4B	Original	1388	4.35	SeedBench 3-2	183/808	2.26	67.71
4B	Native SFT	176	0.55	SeedBench 3-2	143/808	1.19	71.09

Appendix

Table 17: Invalid generated answers over 17 views (31,873 records), excluding paired CROP reordering. Record rate pools invalid outputs; view-mean rate averages their per-view fractions. The repair bound hypothetically makes every invalid answer correct while retaining all valid decisions; it is not an achieved result. All original records remain in the reported accuracy.

External mean (17 views)
Scale	Unadapted	Token-SFT	AgriJev	$\Delta$
0.8B	47.98	54.17	53.55	-0.62
2B	53.61	59.98	59.19	-0.79
4B	67.00	71.70	70.47	-1.23
9B	71.26	75.15	75.44	+0.29
Internal BIO accuracy

Appendix

Table 18: Prior same-inference adaptation and loss controls (%). Unadapted Qwen3.5, full-vocabulary Token-SFT, and AgriJev share candidate scoring. Token-SFT and AgriJev share the agricultural training condition;

\Delta

is AgriJev minus Token-SFT in percentage points.

Supervised encoder controls
Model	Typing	Assertion	Event	BIO accuracy	BIO macro-F1
AgriJev-0.8B	96.17	89.83	80.25	92.75	79.63
AgriJev-2B	96.70	95.46	84.00	94.03	83.77
AgriJev-4B	97.12	95.89	84.15	94.46	84.33
AgriJev-9B	97.62	95.30	81.93	95.57	86.37
GLiClass-FT ( Stepanov et al., 2025 )	96.70	53.51	43.19	93.52	82.75

Appendix

Table 19: Prior candidate-scoring internal evaluation (%). All models train on 41,724 records. Citations identify pretrained families; FT/token-SFT checkpoints are trained in this study. Typing, assertion, and event tests have 3,578, 7,156, and 1,306 views; events use silver labels. BIO is an internal 1,173-record subset with fixed 27-class macro-F1. Bold marks the highest result in each column, including ties.

	0.8B			2B			4B			9B
Dataset	Base	AgriJev	Token	Base	AgriJev	Token	Base	AgriJev	Token	Base	AgriJev	Token
AgriEval	46.16	48.79	49.01	48.67	50.60	51.11	54.50	58.51	57.84	57.96	62.20	62.16
CROP	76.78	84.31	84.02	78.94	86.62	86.60	86.85	91.75	91.69	88.48	91.10	92.11
CMMLU agronomy	39.05	46.15	49.11	49.11	53.85	54.44	61.54	68.64	69.23	73.96	76.33	75.74
TMMLU+ agriculture	24.31	24.31	28.47	32.64	38.19	38.19	43.06	51.39	56.25	53.47	65.97	63.89
C-Eval plant protection	44.72	48.74	48.74	55.28	63.32	60.80	75.88	80.40	80.40	81.91	83.42	81.91

Appendix

Table 20: Prior scoring-only external scale comparison (%). Base denotes unadapted Qwen3.5; Token denotes the matched full-vocabulary answer-token control. All trained decoder rows use the same current release. CROP reordered is paired and excluded from the mean.

Model	p50	p95	Throughput	Backbone
AgriJev-0.8B	42.96	43.62	23.25	BF16
AgriJev-2B	41.90	70.84	21.95	BF16
AgriJev-4B	60.93	66.54	16.33	BF16
AgriJev-9B	93.80	96.53	11.02	BF16
Token-SFT-0.8B	43.44	44.90	22.92	BF16
Token-SFT-2B	44.55	45.00	22.45	BF16

Appendix

Table 21: Prior scoring-contract end-to-end latency for completed supervised models: A100 40GB, batch 1, 32 BIO texts, 27 candidates, two warmups and three repetitions (96 durations/model). p50/p95 are milliseconds; throughput is records/s. Decoder logits/probabilities use FP32 even when backbone inference is BF16. Hosted Jev is excluded from this local table.

Language Modeling

Language modeling develops systems that learn patterns in text and other sequences to understand, predict, and generate language.

See all 519 papers

Fine-Tuning General-Purpose Large Language Models for Agricultural Applications:A Reproducible Framework and Evaluation Protocol Based on Qwen3-8B
Jun 27, 2026Zhaoyang Li, Ruijie Zhang, Jiaqi Liu +1LLM Fine-Tuning
1Sanya University · 2Hebei International Studies University
Evaluating and Benchmarking the System One Model Jev
Sep 29, 2026Tobias Deußer, Lorenz Sparrenberg, Rafet SifaMultilingual Language Model Evaluation LLM Evaluation
University of Bonn, Bonn, Germany · Lamarr-Institute for Machine Learning and Artificial Intelligence, Bonn, Germany · Fraunhofer IAIS, Sankt Augustin, Germany
Specialized Decision Models vs. General-Purpose LLMs: Benchmarking Jev Across Knowledge, Reasoning, and Multilingual Tasks
Oct 8, 2026Xing Li, Qingcheng Chang, Jinzhong Ning +5LLM Evaluation Multiple-Choice Question Answering
Dalian Maritime University · Dalian University of Technology

Jun 27, 2026J/K move · Enter open · S save

Fine-Tuning General-Purpose Large Language Models for Agricultural Applications:A Reproducible Framework and Evaluation Protocol Based on Qwen3-8B

Zhaoyang Li, Ruijie Zhang, Jiaqi Liu, Zhaoji Sun

1Sanya University · 2Hebei International Studies University

LLM Fine-Tuning

Open paper AskPDF arXiv

General-purpose large language models (LLMs) have demonstrated strong abilities in opendomain question answering, information extraction, and text generation. Agricultural applications, however, are domain-specific, region-dependent, time-sensitive, and safety-critical. Without data governance, expert evaluation, and evidence constraints, an agricultural assistant mayproduce unreliable advice on crop diseases, pesticide use, fertilization, or policy interpretation.To avoid presenting unverified simulated numbers as real experimental findings, this paper doesnot report any model-performance claims that have not been produced by an actual training runand expert evaluation. Instead, we propose AgriTune-R, a reproducible and auditable frameworkfor adapting general-purpose LLMs to agricultural tasks. The framework selects the publiclyverifiable Qwen3-8B model as the recommended base model and integrates agricultural datagovernance, instruction construction, LoRA/QLoRA parameter-efficient fine-tuning, retrievalaugmented generation, expert evaluation, and safety control for high-risk questions. The contributions are: (1) a structured workflow for agricultural LLM adaptation; (2) an evaluationprotocol for agricultural knowledge QA, pest and disease consultation, cultivation management,and policy explanation; (3) an expert-review rubric combining factuality, safety, evidence consistency, and uncertainty expression; and (4) a clear separation between protocol design andempirical conclusions, providing an executable baseline for future empirical studies.

Abstract

Language Modeling

Explore similar work

Fine-Tuning General-Purpose Large Language Models for Agricultural Applications:A Reproducible Framework and Evaluation Protocol Based on Qwen3-8B

Evaluating and Benchmarking the System One Model Jev

Specialized Decision Models vs. General-Purpose LLMs: Benchmarking Jev Across Knowledge, Reasoning, and Multilingual Tasks