Jainamshahhh

agrireason-4b

README

License: apache-2.0

Headline result

On a held-out slice of the AgriGround distribution (100 real-farmer questions never seen in training, judged blind and pairwise against the base's answers), AgriReason-4B is preferred 89.0% of the time (tie-allowed) / 92.4% forced-choice (tuned 85 / base 7 / tie 8).

This is the evaluation that matches how the challenge scores entries: on held-out data drawn from the same distribution as the training set. The gap is large because the base model, faced with a terse farmer question, tends to hedge, ask for clarification, or return an encyclopedia entry, whereas AgriReason returns the direct, structured advisory the dataset teaches.

Training details

Base: Qwen/Qwen3.5-4B (text decoder, Qwen3_5ForCausalLM), Apache-2.0.
Method: LoRA SFT, completion-only loss.
LoRA: rank 32, alpha 64, dropout 0.05, target = all linear projections.
Optimizer: AdamW, lr 1e-4, cosine, warmup 0.05, weight decay 0.01, grad-clip 1.0, bf16.
Schedule: 3 epochs, effective batch 16, max sequence length 1792, SDPA attention, A100-80GB.
Data: 19,500 rows, a quality-curated subset of the 26,032-row AgriGround set (all grounded synthetic-reasoning + crop-domain rows, plus the richest, category-balanced real-farmer Q&A). See the dataset card.

Evaluation protocol

Held-out anchor (headline): 100 AgriGround-distribution questions excluded from training, generated by both base and tuned model under identical greedy decoding (1536-token budget), then judged blind: A/B order randomized, model identity hidden, scored accuracy-first with completeness/specificity as tie-breakers. Win rate = (wins + 0.5*ties)/n.
Training was decontaminated against public agronomy benchmarks (no shared 8-grams).

Limitations & responsible use

Confident but unverifiable outputs. The training data includes real farmer-advisory interactions that answered time- and place-specific questions (e.g., local weather, market prices). The model therefore will confidently produce such answers, which it cannot actually know. Do not rely on it for real-time weather, prices, or any live data. Treat all specifics (rates, dates, product names) as suggestions to verify against the current product label and local extension advice.
Not a substitute for local agricultural extension services, a licensed agronomist, or the product label. Crop-safety screening in the training data is best-effort, not exhaustive.
Knowledge is bounded by the base model and the training corpus; not for high-stakes decisions without expert review. This is a research artifact (a LoRA adapter).

Reproducibility & license

Data-build pipeline, eval harness, and blind-judge verdicts are released alongside the model.
License: Apache-2.0 (adapter and base). Training data is governed by its own per-row licenses (see the dataset card; the compilation is CC-BY-NC-SA-4.0).

Usage

python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-4B")
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "Jainamshahhh/agrireason-4b").eval()

msg = [{"role": "user", "content": "My wheat in Punjab during rabi has yellow rust. What should I apply and when?"}]
ids = tok.apply_chat_template(msg, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=512)[0][ids.shape[1]:], skip_special_tokens=True))

Author: Jainam Shah. Built with Adaptive Data (Adaption Labs) and the open Qwen model.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Jainamshahhh

Model Tree

Base

Qwen/Qwen3.5-4B

Adapter

this model

Input Modalities

Text

Image

Video

Output Modalities