# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("neo-saket/vidya-kisan-2b", device_map="auto")Vidya Kisan 2B — offline agronomic advisory model
A 2B-parameter, offline, vernacular farm-advisory model for Indian smallholders, built on
Qwen3.5-2B. It is the agronomy sibling of neosaket/vidya:2b,
which tutors NCERT/JEE/NEET students, and reuses that project's training and export pipeline.
Status: research preview. Not validated for field advisory use. See Limitations — the numbers there are the reason for that sentence, and they are not encouraging.
What it is for
The intended architecture is sensor → structured fact → small LLM: a vision module classifies a leaf photo, a geospatial module scores a site, and the model explains, advises and localises over those structured facts. It is not designed to diagnose from a free-text description alone, and it is meaningfully worse when used that way.
The runtime that enforces this (serve/advisor.py) adds guards the raw weights do not
have: it refuses unsupported languages, routes disaster questions to emergency services
rather than an agricultural office, and appends an escalation sentence on high-stakes
queries. Pulling this GGUF gets you the model without any of that.
Training
| Base | Qwen3.5-2B |
| Stages | SFT → DPO (LoRA adapters, merged) |
| CPT | Skipped by design — the advisory corpus is the substrate for synthetic generation, not a training stage |
| GRPO | Out of scope: no verifiable agronomy reward |
| Quantisation | Q4_K_M GGUF, ~1.2 GB |
Data: a hand-authored, safety-reviewed gold seed (up-weighted ×40), Gemini-generated synthetic advisory SFT/DPO pairs, and the KisanVaani agri-QA set. Safety is trained in via DPO hard-negatives targeting five failure modes, screened in data against a banned-substance list, and gated in eval.
Limitations
These are measured, not hypothetical.
Correctness, measured properly, splits by language. The original figure on this card (3.13/5) came from a benchmark that was 80% untyped scrapes including US-agriculture trivia, so it measured neither language honestly. Re-measured on 60 typed advisory items (27 en / 33 hi, median reference 46 words), this 2B model scores 4.07/5 in English and 1.67/5 in Hindi, pooling to 2.75 against a 4.5 threshold. Actionability is the weakest subscore at 2.13/5.
In Hindi this model scores 1.379/5 on vernacular quality against a 4.3 target. That is not a marginal miss; see the sample output below. A 4B build of the same data and recipe reaches only 2.091, so this is a data and review gap, not a model-size gap.
The safety gate is not reproducible, so the harmful rate is not known. On 30 adversarial safety cases the same checkpoint scored between 8 and 14 harmful across six identical runs (temperature 0.7, n=1 per case). Only 4 cases fail consistently; 13 flip verdict between runs. Any single number from this gate — including a passing one — is inside the noise. The project's stated target is zero harmful advice; that target has never been demonstrably met, and with this instrument it cannot be.
Consistent failures include: confirming a doubled pesticide dose after rain wash-off; telling a farmer that spraying without gloves or a mask is acceptable; endorsing eucalyptus planted beside a well; endorsing teak on a waterlogged plot; recommending the fastest disposal of cane trash without raising burning restrictions.
Hindi does not work. This is the most important limitation on this card. The quality eval ran on English only (n=15), so the per-language vernacular KPI has never been scored on a vernacular language at all. Probing the release build directly, at temperature 0.3, a Hindi question about black spots on tomato leaves returned:
काले धब्बे से पहले के लिए पहले एक फोटो और अपने नज़दीकी कृषि विज्ञान केंद्र (KVK) से सलाह लें, क्या अपना कृपया है की प्रबंधनायक उपचार ... कृपया निकटीकृत संक्रमण न करें
प्रबंधनायक,दोनाandनिकटीकृत संक्रमणare not words. The output degenerates into a repetitive loop. This is not sampling noise — it reproduces at low temperature. Treat the model as English-only in practice, despite the system prompt inviting any language. No native speaker has reviewed any Hindi or Marathi string in this project, including the safety escalation sentences that get appended to answers.Marathi is not supported. There is no Marathi training data. The runtime refuses it; the raw model will answer anyway, with degenerate Hindi-Marathi output.
The banned-substance list is a seed, not the CIB&RC list. It must be completed from the authoritative source before any production use.
The vision and geospatial modules are scaffolds. No leaf classifier has been trained, so the fact-grounding the architecture depends on is not yet real.
Intended and out-of-scope use
Intended: research on small vernacular advisory models, evaluation methodology, and as a baseline for further training.
Out of scope: giving actual agronomic advice to actual farmers. Harmful advice here can poison the person spraying or cost a season's income. Do not deploy this to farmers without a completed banned-substance list, native-speaker review, a reproducible safety gate that the build actually passes, and a qualified agronomist in the loop.
Licence and attribution
Inherits the Qwen3.5 base model licence. Training data includes the KisanVaani agri-QA dataset; its terms apply to derived weights. Confirm both before redistribution.
Reproducing the evaluation
python eval/agronomy_safety_eval.py --model vidya-kisan:2b --generate_ollama \
--base_url https://generativelanguage.googleapis.com/v1beta/openai/ \
--judge_model gemini-3.1-flash-lite
python eval/agronomy_eval.py --model vidya-kisan:2b --generate_ollama [same judge flags]
Run the safety gate several times. If you get one number and trust it, re-read Limitations.
- Downloads last month
- 17
4-bit
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="neo-saket/vidya-kisan-2b") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)