Text Generation
Transformers
Safetensors
English
Hindi
Spanish
qwen3_5
image-text-to-text
router
planner
orchestration
tool-use
structured-output
json
lora
qwen3.5
health-assistant
conversational
Instructions to use meet4150/xevyo-01 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use meet4150/xevyo-01 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="meet4150/xevyo-01") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("meet4150/xevyo-01") model = AutoModelForMultimodalLM.from_pretrained("meet4150/xevyo-01", device_map="auto") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use meet4150/xevyo-01 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "meet4150/xevyo-01" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "meet4150/xevyo-01", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/meet4150/xevyo-01
- SGLang
How to use meet4150/xevyo-01 with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "meet4150/xevyo-01" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "meet4150/xevyo-01", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "meet4150/xevyo-01" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "meet4150/xevyo-01", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use meet4150/xevyo-01 with Docker Model Runner:
docker model run hf.co/meet4150/xevyo-01
File size: 8,941 Bytes
a46f204 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 | """Ask one of OUR endpoints for a turn plan on every gold row, then score it.
python3 eval/run_eval.py --base http://127.0.0.1:18220 --model xevyo-01 \
--gold 'eval/gold/gold_*.jsonl' --out /data/xevyo-01/eval/xevyo-01_r1 --concurrency 8
Every request is what serving will send: the student's own prompt,
thinking off, temperature 0, and the per-turn schema through serve/vllm_schema.py
(response_format json_schema, the method that works on vLLM 0.17). Each answer is parsed,
checked with spec/validate.py and scored with eval/score.py.
Baselines through the same script:
* base Qwen3.5-9B zero-shot: --base <port serving the base> --model <name>
* a larger prompted model: --base <url serving it> --model <name>
--system-prefix-file <rules.md>
(the file is put in front of the student's system prompt, so the larger model
knows the rules the labels follow; the student never sees it)
Latency: --latency-pass N sends the first N rows again one at a time (concurrency 1)
and records their wall time; the scored latency comes from that pass when given.
The gold rows carry many different offered-tool sets (448 schemas in 1,610 rows, 93 in
the first 200), and vLLM 0.17 compiles a schema it has not cached on the request path, so that pass includes compile time on about half its
rows. --latency-same-tools N adds a second sequential pass over N rows that all share
the most common schema (335 gold rows have it; compiled once): the steady state of a
user whose tools do not change. Its p50 / p90 go into the report's meta.
Point this at your own xevyo-01 endpoint, never at a live service's model.
Standard library + jsonschema.
"""
import argparse
import json
import sys
import threading
import time
import urllib.error
import urllib.request
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / "spec"))
sys.path.insert(0, str(ROOT / "serve"))
sys.path.insert(0, str(ROOT / "eval"))
import render
import score
import validate
from vllm_schema import response_format
def post(base, body, timeout):
req = urllib.request.Request(base.rstrip("/") + "/v1/chat/completions", data=json.dumps(body).encode(),
headers={"Content-Type": "application/json"})
with urllib.request.urlopen(req, timeout=timeout) as r:
return json.loads(r.read())
def ask(args, row, prefix):
example = row["input"]
# Round 4: the context window; a row's own "window" wins over --window.
messages = render.render_messages(example, row.get("window") or args.window)
if prefix:
messages[0] = {"role": "system", "content": prefix + "\n\n---\n\n" + messages[0]["content"]}
body = {"model": args.model, "messages": messages, "temperature": 0, "max_tokens": args.max_tokens,
"response_format": response_format(example["offered_tools"]),
"chat_template_kwargs": {"enable_thinking": False}}
out = {"id": row["id"], "plan": None, "raw": None, "errors": [], "latency_s": None}
for attempt in range(3):
t0 = time.time()
try:
resp = post(args.base, body, args.timeout)
except (urllib.error.URLError, TimeoutError, ConnectionError) as e:
out["errors"] = [f"http: {e}"]
time.sleep(2 * (attempt + 1))
continue
out["latency_s"] = time.time() - t0
text = resp["choices"][0]["message"].get("content") or ""
out["raw"] = text
out["usage"] = resp.get("usage")
try:
plan = json.loads(text)
except json.JSONDecodeError as e:
out["errors"] = [f"json: {e}"]
return out
out["plan"] = plan
out["errors"] = validate.validate_plan(plan, example)
return out
return out
def main(argv=None):
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--base", required=True, help="e.g. http://127.0.0.1:18220 (your xevyo-01 endpoint)")
ap.add_argument("--model", required=True)
ap.add_argument("--gold", nargs="+", required=True)
ap.add_argument("--out", required=True, help="output prefix (directory + name)")
ap.add_argument("--concurrency", type=int, default=8)
ap.add_argument("--max-tokens", type=int, default=512)
ap.add_argument("--timeout", type=float, default=120)
ap.add_argument("--limit", type=int, default=0)
ap.add_argument("--system-prefix-file", default="")
ap.add_argument("--latency-pass", type=int, default=0)
ap.add_argument("--latency-same-tools", type=int, default=0,
help="second sequential pass over N rows that share the most common offered-tool set")
ap.add_argument("--title", default="")
ap.add_argument("--window", default="trained", choices=render.WINDOWS,
help="context window the inputs are rendered in (the backend's XEVYO01_CONTEXT); default trained")
ap.add_argument("--no-score", action="store_true", help="only write predictions (rows without a target)")
args = ap.parse_args(argv)
port = args.base.rstrip("/").rsplit(":", 1)[-1]
if not (port.isdigit() and 18200 <= int(port) <= 18299):
sys.exit(f"refusing {args.base}: only our own ports 18200-18299")
gold = score.load_jsonl(args.gold)
if args.limit:
gold = gold[:args.limit]
prefix = Path(args.system_prefix_file).read_text(encoding="utf-8") if args.system_prefix_file else ""
out = Path(args.out)
out.parent.mkdir(parents=True, exist_ok=True)
pred_path = Path(str(out) + "_predictions.jsonl")
lock = threading.Lock()
done = [0]
t_start = time.time()
def job(row):
res = ask(args, row, prefix)
with lock:
done[0] += 1
if done[0] % 100 == 0:
print(f"{done[0]}/{len(gold)} in {time.time() - t_start:.0f}s", flush=True)
return res
with ThreadPoolExecutor(args.concurrency) as pool:
preds = list(pool.map(job, gold))
wall = time.time() - t_start
if args.latency_pass:
seq = {}
for row in gold[:args.latency_pass]:
r = ask(args, row, prefix)
seq[row["id"]] = r["latency_s"]
for p in preds:
p["latency_s"] = seq.get(p["id"])
same_tools = None
if args.latency_same_tools:
# Rows are grouped by the schema that is sent (what the server's grammar cache keys on),
# not by the raw tool list: lists that differ only in tools a plan cannot call share one.
key = lambda row: json.dumps(response_format(row["input"]["offered_tools"]), sort_keys=True)
counts = {}
for row in gold:
counts[key(row)] = counts.get(key(row), 0) + 1
common = max(counts, key=counts.get)
rows_same = [row for row in gold if key(row) == common][:args.latency_same_tools]
lat = sorted(x for x in (ask(args, row, prefix)["latency_s"] for row in rows_same) if x is not None)
if lat:
same_tools = {"rows": len(lat), "p50_s": round(lat[len(lat) // 2], 4),
"p90_s": round(lat[min(len(lat) - 1, int(len(lat) * 0.9))], 4),
"mean_s": round(sum(lat) / len(lat), 4)}
with open(pred_path, "w", encoding="utf-8") as f:
for p in preds:
f.write(json.dumps(p, ensure_ascii=False) + "\n")
if args.no_score:
print(f"wrote {pred_path} ({len(preds)} rows, {wall:.0f}s)")
return 0
report, rows, missing = score.score(gold, preds)
meta = {"base": args.base, "model": args.model, "gold": args.gold, "rows": len(gold),
"concurrency": args.concurrency, "wall_s": round(wall, 1),
"system_prefix_file": args.system_prefix_file or None,
"latency": f"sequential pass over {args.latency_pass} rows" if args.latency_pass else f"under concurrency {args.concurrency}",
"latency_same_tools": same_tools,
"unparsed": sum(1 for p in preds if p["plan"] is None),
"invalid": sum(1 for p in preds if p["plan"] is not None and p["errors"])}
traffic = score.traffic_weighted(rows) # TRAFFIC_WEIGHTS mix, see eval/score.py
Path(str(out) + ".json").write_text(json.dumps({"meta": meta, "report": report, "traffic": traffic}, indent=1),
encoding="utf-8")
with open(str(out) + "_rows.jsonl", "w", encoding="utf-8") as f:
for r in rows:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
md = score.markdown(report, args.title or f"{args.model} on {len(gold)} gold rows")
md += "\n" + score.traffic_markdown(traffic)
md += f"\n{json.dumps(meta)}\n"
Path(str(out) + ".md").write_text(md, encoding="utf-8")
print(md)
return 0
if __name__ == "__main__":
sys.exit(main())
|