FieldEvidence — Llama 4 Scout Agriculture Adapter
FieldEvidence is an evidence-grounded agricultural decision-support adapter trained with Adaption AutoScientist. It is designed to answer agricultural questions directly while stating the boundary of the available evidence and identifying what requires local verification.
Results
| Evaluation | Frozen base | FieldEvidence | Change |
|---|
| Project dataset | 33.1% | 66.9% | +33.8 points |
| Held-out Agriculture category | 26.0% | 74.0% | +48.0 points |
The held-out Agriculture evaluation used 100 unseen category samples and produced 20 base-model wins, 68 adapted-model wins, and 12 ties. The platform reported 26% versus 74% after accounting for ties. The judge recorded by the training backend was google/gemini-3.1-pro-preview through Adaption's in-house evaluation pipeline.
Two additional Llama 4 Scout training lineages reached 69% and 68% on the Agriculture evaluation. Across the three runs, the adapted-model mean was 70.3%, compared with a 30.0% mean baseline. These follow-up runs are supporting replication evidence; the 74% run is the released champion.
Training data
The 3,149-record research core combines:
- 1,199 original field-diagnostic reasoning scenarios with agronomic-validity filtering;
- 580 source-grounded irrigation records declared under MIT licensing; and
- 1,370 KisanVaani Agriculture Q&A records declared under Apache-2.0.
Each record retains a task type, source, license, evidence identifier, verification status, prompt, and completion. A separate 554-fact-group private research holdout was excluded from training, with zero fact-group overlap. The public training core contains no exact duplicate prompts.
Adaption Adaptive Data improved the displayed quality score from 7.0 to 8.5. AutoScientist then created a 20,024-row mixture from the research core, 8,900 additional domain records, and 8,000 general-purpose records. After filtering and splitting, it contained 19,677 train, 175 validation, and 172 test examples.
Public dataset:
Training recipe
- Base model used by the exported adapter:
togethercomputer/Llama-4-Scout-17B-16E-Instruct_bnb_4bit
- Model family:
meta-llama/Llama-4-Scout-17B-16E-Instruct
- Method: supervised fine-tuning with LoRA
- Epochs: 3
- Rank: 64
- Alpha: 128
- Dropout: 0
- Learning rate: 1e-4
- Scheduler: cosine
- Warmup ratio: 0.1
- Minimum learning-rate ratio: 0.1
- Weight decay: 0.01
- Gradient clipping: 1.0
- Train on input tokens: false
- Target modules:
q_proj,k_proj,v_proj,o_proj,shared_expert.gate_proj,shared_expert.up_proj,shared_expert.down_proj,feed_forward.gate_proj,feed_forward.up_proj,feed_forward.down_proj
Intended use
FieldEvidence is intended for source-grounded Agriculture Q&A, cautious next-step guidance for crop and field observations, agricultural education, extension-support prototypes, and research into evidence-boundary behavior in domain-adapted language models.
The preferred response structure is:
1Answer: [best supported answer or next step]
2Boundary: [what the evidence cannot confirm and what requires local verification]
Limitations and safety
FieldEvidence is not a field diagnosis, pesticide label, laboratory test, live weather service, regulatory authority, or replacement for a qualified agronomist or extension adviser.
The model may produce incorrect, outdated, geographically inappropriate, or overly general guidance. It should not select or dose pesticides, fertilizers, veterinary products, or other consequential treatments without current local verification. Historical and regional source records must not be interpreted as present field conditions.
Confirm consequential actions using current local extension guidance, direct field observations, appropriate sampling or testing, product labels, and qualified professionals.
Release notes
This repository contains the PEFT/LoRA adapter produced from the champion AutoScientist run. Use it with the compatible base-model revision named in adapter_config.json. Use of the base model remains subject to Meta's applicable Llama license and acceptable-use terms. Dataset components retain their respective source licenses.
Acknowledgements
Built for the AutoScientist Challenge using Adaption Adaptive Data and AutoScientist. Source-backed components include the KisanVaani Agriculture Q&A collection and agricultural extension passages represented in the released dataset provenance fields.
