GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
WIZARDIAN
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
WIZARDIAN
Model Tree
Qwen/Qwen2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
The model has two narrow jobs:
The model is never asked to recall weather, air quality or places. Every number, time and proper noun in its answer is checked against the fact sheet by a grounding verifier; a draft that fails is regenerated once with feedback, and if it fails again the system falls back to a deterministic template.
Parse (system prompt, few-shot, temperature 0, JSON schema grammar):
text
You extract structured fields from a question about an outdoor activity in India. activity must be one of ['running', 'cycling', 'golf', 'cricket', 'football', 'tennis', 'hiking', 'walking', 'picnic', 'event', 'photography', 'yoga', 'birdwatching', 'kite', 'stargazing'], or "out_of_scope" if the question is not about going outdoors, or "unknown" if unclear. location is the place name copied exactly from the question, or an empty string if none is written. radius_km is the distance the question states, or 0. profiles lists any of ['asthma', 'senior', 'child', 'pregnancy', 'heart'] that the question mentions (senior = elderly or parents, child = kids). Reply with JSON only.Compose (system prompt, temperature 0.3, max 320 tokens):
text
You are a friendly outdoor-activity advisor for India. Write the answer using ONLY the facts in the FACT SHEET. Never add a number, place, time or weather detail that is not in the fact sheet, and copy place names and times exactly. Write 80-150 words of plain text (no headings, no tables): start with a one-sentence recommendation naming the place and time, then explain why in two or three short sentences using the strongest reasons, mention the main thing to avoid, and end with one practical tip. If the fact sheet says the activity is not advised, say so clearly first. Write flowing sentences: no labels such as "Tip:", and never copy the fact sheet's field names. Describe air quality only with the category the fact sheet gives (Good, Satisfactory, Moderate, Poor, Very Poor, Severe).The exact JSON schema and generation settings are in geoai_config.json in this repository.
python
from transformers import AutoModelForCausalLM, AutoTokenizerrepo = "WIZARDIAN/Bhugoole-geoai-outdoor-slm"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto")The LoRA adapter alone, on top of the base model:
python
from peft import PeftModelbase = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")model = PeftModel.from_pretrained(base, "WIZARDIAN/Bhugoole-geoai-outdoor-slm", subfolder="adapter")With llama.cpp (what the API uses):
python
from llama_cpp import Llamallm = Llama.from_pretrained("WIZARDIAN/Bhugoole-geoai-outdoor-slm", "bhugoole-geoai-outdoor-slm-q4_k_m.gguf")QLoRA fine-tune of Qwen/Qwen2.5-1.5B-Instruct for slot extraction (question -> strict JSON).
finetune/make_parse_dataset.py (labels known by construction; typos, lower case, Hinglish fillers,
missing places and out-of-scope requests are over-sampled).Scores below come from this script (transformers, 4-bit, greedy decoding); see the end of this file for the live API's path (llama.cpp, JSON-schema decoding).
| Metric | Synthetic held-out (120): base | fine-tuned | Real questions (30): base | fine-tuned | Unseen test (40): base | fine-tuned |
|---|---|---|---|---|---|---|
| exact | 50.8% | 95.8% | 73.3% | 100.0% | 75.0% | 97.5% |
| activity | 66.7% | 95.8% | 93.3% | 100.0% | 87.5% | 100.0% |
| location | 76.7% | 100.0% | 86.7% | 100.0% | 95.0% | 97.5% |
| profiles | 85.0% | 100.0% | 86.7% | 100.0% | 90.0% | 100.0% |
| valid json | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
The synthetic held-out set comes from the same templates as the training data, so it is optimistic.
finetune/real_eval.jsonl (30) and finetune/test_unseen.jsonl (40) are hand-written and never used for
training; the unseen set was written before this version was trained, with places and phrasings the generator
does not use.
Measured with finetune/eval_parse.py. Version 1 was trained on 2,700 examples from the first generator; the error
analysis of version 1 on the 30 real questions showed two patterns (off-topic requests that name a place, and
places after unusual connectors such as "trek to sinhagad"), which version 2's generator covers in general form.
Because the 30 questions guided that analysis, the unseen set is the fairer measure.
| Metric | Base Qwen2.5-1.5B | Version 1 | Version 2 (deployed) |
|---|---|---|---|
| real questions (30), exact | 76.7% | 93.3% | 100.0% |
| unseen test (40), exact | 72.5% | 77.5% | 97.5% |
| unseen test (40), activity | 85.0% | 85.0% | 100.0% |
| unseen test (40), location | 90.0% | 80.0% | 97.5% |
| unseen test (40), profiles | 92.5% | 100.0% | 100.0% |
| off-topic requests caught (6 in the unseen set) | 1 | 1 | 6 |
| answers written by the SLM and passing the grounding check (8 live questions) | 7/8 | 8/8 | 7/8 |
The only miss of version 2 on the 70 hand-written questions: "Worli sea face" was read as "Worli sea". Answer writing was not a training target; the 7/8 versus 8/8 difference is within run-to-run variation, and a draft that fails the check is replaced by the verified template.
The running assistant uses only this open model; no hosted language model is called.
You extract structured fields from a question about an outdoor activity in India. activity must be one of ['running', 'cycling', 'golf', 'cricket', 'football', 'tennis', 'hiking', 'walking', 'picnic', 'event', 'photography', 'yoga', 'birdwatching', 'kite', 'stargazing'], or "out_of_scope" if the question is not about going outdoors, or "unknown" if unclear. location is the place name copied exactly from the question, or an empty string if none is written. radius_km is the distance the question states, or 0. profiles lists any of ['asthma', 'senior', 'child', 'pregnancy', 'heart'] that the question mentions (senior = elderly or parents, child = kids). Reply with JSON only.You are a friendly outdoor-activity advisor for India. Write the answer using ONLY the facts in the FACT SHEET. Never add a number, place, time or weather detail that is not in the fact sheet, and copy place names and times exactly. Write 80-150 words of plain text (no headings, no tables): start with a one-sentence recommendation naming the place and time, then explain why in two or three short sentences using the strongest reasons, mention the main thing to avoid, and end with one practical tip. If the fact sheet says the activity is not advised, say so clearly first. Write flowing sentences: no labels such as "Tip:", and never copy the fact sheet's field names. Describe air quality only with the category the fact sheet gives (Good, Satisfactory, Moderate, Poor, Very Poor, Severe).from transformers import AutoModelForCausalLM, AutoTokenizerrepo = "WIZARDIAN/Bhugoole-geoai-outdoor-slm"tok = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto")from peft import PeftModelbase = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")model = PeftModel.from_pretrained(base, "WIZARDIAN/Bhugoole-geoai-outdoor-slm", subfolder="adapter")from llama_cpp import Llamallm = Llama.from_pretrained("WIZARDIAN/Bhugoole-geoai-outdoor-slm", "bhugoole-geoai-outdoor-slm-q4_k_m.gguf")