rajd9621/drishti-slm

Drishti SLM

A small open model, fine-tuned to reason over fetched geospatial and weather data for an outdoor-activity advisor in India. Built for the GeoAI Hackathon @ CSRE, IIT Bombay.

The model never recalls weather or geography. It is handed a set of places, hours and the real numbers for each, and it chooses between them.

What it does

Three jobs, one adapter:

job	input	output
parse	"I have mild asthma, where's good for a morning walk near Powai in the next two days?"	strict JSON slots: activity, place, radius, time, persona
decide	a FACTS block: the scored (place, hour) surface with real weather, AQI, travel time and geospatial flags	which place and hour, why, why the runner-up lost, what to carry, what to watch for
narrate	the chosen recommendation plus its evidence	a short, grounded answer that names its sources

The decide job is the reason this model exists. A deterministic engine does the fetching and the scoring; the model makes the judgement call that a scorer cannot: that a shaded place beside water is worth more on a hot afternoon, that a ground downwind of an expressway is worth less at 8 AM than at 6 PM, that a venue shut at the suggested hour is not a choice at all, and that a venue forty minutes away only pays off if the conditions really are better.

Grounding, by construction

Every decision is validated before it reaches the user (app/planner.py):

• a best_place that is not in the supplied FACTS is rejected;
• a best_hour that the place does not have falls back to that place's best hour;
• any sentence quoting a number the facts do not contain is dropped;
• carry and watch items outside a fixed vocabulary are discarded;
• if the decision does not survive, the engine's own pick stands.

The model can therefore only ever choose between real, already-fetched options. It cannot introduce a place, an hour or a number of its own.

Base model

Qwen/Qwen2.5-1.5B-Instruct - 1.54B parameters, Apache-2.0, open weights. Other options, all under the 3B limit: Qwen/Qwen2.5-0.5B-Instruct, microsoft/Phi-3-mini-4k-instruct, meta-llama/Llama-3.2-1B-Instruct. Swap with --base in training/train_lora.py.

Training

QLoRA (4-bit NF4), LoRA rank 16 / alpha 32 on all attention and MLP projections. Two epochs, cosine schedule, lr 2e-4, effective batch 16, max length 1536.

pip install -r training/requirements-train.txt
python training/generate_dataset.py --out data --n 1500
python training/train_lora.py --data data/all.jsonl --out out/drishti-lora

Data is synthetic: ~1500 chat-format examples covering the three jobs, drawn from realistic Indian weather, AQI, travel-time and geospatial ranges. It teaches the shape of the task - choose only from what you were given, quote no number you were not given, keep it short - not the climate of any city. Regenerate with a different --seed for a fresh split.

Using it

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-1.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = PeftModel.from_pretrained(
    AutoModelForCausalLM.from_pretrained(base), "<this-repo>")

msgs = [{"role": "system", "content": DECIDE_SYSTEM},
        {"role": "user", "content": "FACTS: " + facts_json}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
print(tok.decode(model.generate(ids, max_new_tokens=320)[0][ids.shape[-1]:],
                 skip_special_tokens=True))

In the assistant itself, set DRISHTI_ENABLE_SLM=1 and point DRISHTI_MODEL_ID at this repo (or at a local path). A quantised GGUF file can be used instead with DRISHTI_SLM_BACKEND=llamacpp and DRISHTI_GGUF_PATH, which is 3-4x faster on CPU.

Limitations

• Synthetic training data. The model has not seen real advisories; it has seen the structure of them. Real behaviour depends on the live data the engine feeds it.
• It decides, it does not compute. Scores, weather values, travel times and distances all come from the engine. The model's contribution is the choice and the framing.
• English. Prompts are English. The assistant's own reply localisation (English, Hindi, Marathi, Tamil, Bengali) is a separate phrase-table layer, not the model.
• Small. At 1.5B it can be inconsistent on unusual phrasings; the deterministic fallback exists precisely for that.
• Not a safety system. Health advice is bounded by fixed persona rules in the engine, not by the model.

Licence

Apache-2.0, matching the base model. Built with open data: Open-Meteo, OpenStreetMap via Overpass, Nominatim, OSRM. Map data © OpenStreetMap contributors.