Safetensors
GGUF
English
Russian
lfm3
reasoning
test-time-compute
inference-time-scaling
code-generation
llama.cpp
negative-results
conversational
Instructions to use HelioAI/LFM3-1.7B-inference-modules-experiments with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use HelioAI/LFM3-1.7B-inference-modules-experiments with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M # Run inference directly in the terminal: llama cli -hf HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M # Run inference directly in the terminal: llama cli -hf HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M # Run inference directly in the terminal: ./build/bin/llama cli -hf HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
Use Docker
docker model run hf.co/HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
- Llama new
- LM Studio
- Jan
- Ollama
How to use HelioAI/LFM3-1.7B-inference-modules-experiments with Ollama:
ollama run hf.co/HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
- Unsloth Desktop
- Pi
How to use HelioAI/LFM3-1.7B-inference-modules-experiments with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:9931/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use HelioAI/LFM3-1.7B-inference-modules-experiments with Docker Model Runner:
docker model run hf.co/HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
- Lemonade
How to use HelioAI/LFM3-1.7B-inference-modules-experiments with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
Run and chat with the model
lemonade run user.LFM3-1.7B-inference-modules-experiments-Q5_K_M
List all available models
lemonade list
- Hermes Agent
How to use HelioAI/LFM3-1.7B-inference-modules-experiments with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:9931/v1 hermes config set model.default HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use HelioAI/LFM3-1.7B-inference-modules-experiments with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:9931/v1 \ --custom-model-id "HelioAI/LFM3-1.7B-inference-modules-experiments:Q5_K_M" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
File size: 8,782 Bytes
67ebcef | 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 | """Apply the deliberation module to already generated reasoning (exactly what it would do online).
The module never changes the reasoning before it acts, and its probes are greedy (deterministic given the prefix),
so running it over recorded trajectories reproduces its online behaviour on those same trajectories. This gives a
fully paired comparison with the recorded baseline.
Module (settings chosen on dev data only):
* every --every tokens of thinking: close at the last sentence end and read a short forced answer
* commit when --agree successive answers are identical
* if the reasoning hit the token limit: answer within the budget instead of being cut off
(short-answer tasks: forced short answer at the limit; code / instruction tasks: thinking closed
--free-reserve tokens before the limit, answer generated with the normal sampling settings)
Output: results/<out_tag>/*.jsonl in gen.py format (+ "module": commit/fallback/none) -> grade.py, compare.py
usage: posthoc_coc.py --src results/orig/main.jsonl results/orig/aime.jsonl --out-tag orig_dm
"""
import argparse
import json
import os
import re
import sys
import threading
import time
import urllib.request
import zlib
from concurrent.futures import ThreadPoolExecutor
from transformers import AutoTokenizer
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from grade import last_boxed, load_suite # noqa: E402
ap = argparse.ArgumentParser()
ap.add_argument("--src", nargs="+", required=True)
ap.add_argument("--out-tag", required=True)
ap.add_argument("--max-tokens", type=int, default=16384)
ap.add_argument("--every", type=int, default=2048)
ap.add_argument("--agree", type=int, default=3)
ap.add_argument("--free-reserve", type=int, default=1536)
ap.add_argument("--commit-tasks", nargs="*", default=["mmlupro", "bbh"],
help="tasks where early commitment is allowed (dev: it hurts numeric math at a 16k budget)")
ap.add_argument("--only-tasks", nargs="*", default=None, help="process only these tasks (others copied unchanged)")
ap.add_argument("--workers", type=int, default=32)
ap.add_argument("--port", type=int, default=8080)
ap.add_argument("--tokenizer", default="models/LFM3-1.7B")
ap.add_argument("--items", default=None, help="glob of item files (default: the eval suite)")
args = ap.parse_args()
THINK_END = 124902
URL = f"http://localhost:{args.port}/completion"
tok = AutoTokenizer.from_pretrained(args.tokenizer)
ENDS = set()
for i_ in range(len(tok)):
t_ = tok.convert_ids_to_tokens(i_)
if t_ is not None and (t_.endswith(".") or t_.endswith("\u010a")):
ENDS.add(i_)
MATH = ("The final answer is $\\boxed{", 48, ["}$", "}\n"], "}$")
SHORT = {"math500": MATH, "aime": MATH, "gsm8k": MATH, "math": MATH,
"mmlupro": ("The answer is (", 3, [")"], ")."), "bbh": ("The answer is ", 24, ["\n", ".\n"], ".")}
SHORT_IDS = {k: tok(v[0], add_special_tokens=False)["input_ids"] for k, v in SHORT.items()}
items = load_suite(args.items) if args.items else load_suite()
lock = threading.Lock()
stats = {"n": 0, "commit": 0, "fallback": 0, "t0": time.time()}
def post(body):
for attempt in range(6):
try:
req = urllib.request.Request(URL, data=json.dumps(body).encode(), headers={"Content-Type": "application/json"})
return json.loads(urllib.request.urlopen(req, timeout=3600).read())
except urllib.error.HTTPError as e:
if e.code == 400:
return None
time.sleep(3)
except Exception:
time.sleep(3)
return None
def extract(task, text):
if task == "mmlupro":
m = re.findall(r"answer is:?\s*\(?\**\s*([A-J])\b", text)
a = m[-1] if m else None
elif task == "bbh":
m = re.findall(r"answer is:?\s*(.+)", text, re.I)
a = m[-1] if m else (last_boxed(text) or "")
a = re.sub(r"\\boxed\{(.*)\}", r"\1", a).strip().strip("*$ .").lower() or None
else:
b = last_boxed(text)
a = re.sub(r"\s+", "", b) if b else None
if a is None or "..." in a or a in ("?", "") or len(a) > 80:
return None
return a
def snap(gen, k, lo=0):
for j in range(min(k, len(gen)), lo, -1):
if gen[j - 1] in ENDS:
return j
return None
def short_probe(ids, gen, cut, task):
suf, n_new, stops, close = SHORT[task]
pr = post({"prompt": ids + gen[:cut] + [THINK_END] + SHORT_IDS[task], "n_predict": n_new, "temperature": 0.0,
"cache_prompt": False, "return_tokens": True, "stop": stops})
if pr is None:
return None, None, 0
c = pr["content"]
if close == "}$":
depth = 1
for j, ch in enumerate(c):
depth += (ch == "{") - (ch == "}")
if depth == 0:
c = c[:j]
break
c = c.strip().rstrip(".").strip() if close == "." else c.strip()
text = suf + c + close
return text, extract(task, text), len(SHORT_IDS[task]) + len(pr["tokens"])
def run(rec):
it = items[rec["id"]]
task = it["task"]
ids = tok.apply_chat_template([{"role": "user", "content": it["prompt"]}], tokenize=True,
add_generation_prompt=True, return_dict=False)
body = rec["completion"][len("<think>"):]
thinking = body.split("</think>", 1)[0]
gen = tok(thinking, add_special_tokens=False)["input_ids"]
out = dict(rec)
out["module"] = "none"
if args.only_tasks is not None and task not in args.only_tasks:
return out
if task in SHORT and task in args.commit_tasks:
prev, answers = 0, []
for k in range(args.every, len(gen), args.every):
cut = snap(gen, k, prev)
if cut is None:
continue
prev = cut
text, ans, n = short_probe(ids, gen, cut, task)
answers.append(ans)
tail = answers[-args.agree:]
if len(tail) == args.agree and tail[0] is not None and all(a == tail[0] for a in tail):
out.update(completion="<think>" + tok.decode(gen[:cut]) + "</think>" + text, n_tokens=cut + n,
truncated=False, module="commit", module_cut=cut)
break
if out["module"] == "none" and rec.get("truncated"):
if task in SHORT:
cut = snap(gen, args.max_tokens - 64)
if cut:
text, ans, n = short_probe(ids, gen, cut, task)
if text is not None:
out.update(completion="<think>" + tok.decode(gen[:cut]) + "</think>" + text, n_tokens=cut + n,
truncated=False, module="fallback", module_cut=cut)
else:
cut = snap(gen, args.max_tokens - args.free_reserve)
if cut:
seed = zlib.crc32(rec["id"].encode()) % 1000003 * 7 + rec.get("sample", 0) + 17
pr = post({"prompt": ids + gen[:cut] + [THINK_END], "n_predict": args.free_reserve,
"temperature": 0.6, "top_p": 0.95, "top_k": 50, "min_p": 0.0, "seed": seed,
"cache_prompt": False, "samplers": ["top_k", "top_p", "temperature"], "return_tokens": True})
if pr is not None:
out.update(completion="<think>" + tok.decode(gen[:cut]) + "</think>" + pr["content"],
n_tokens=cut + len(pr["tokens"]), truncated=pr.get("stop_type") == "limit",
module="fallback", module_cut=cut)
with lock:
stats["n"] += 1
stats["commit"] += out["module"] == "commit"
stats["fallback"] += out["module"] == "fallback"
if stats["n"] % 250 == 0:
print(f"{stats['n']} done commits {stats['commit']} fallbacks {stats['fallback']} "
f"{(time.time() - stats['t0']) / 60:.1f} min", flush=True)
return out
for src in args.src:
recs = [json.loads(l) for l in open(src, encoding="utf-8")]
dst = os.path.join("results", args.out_tag, os.path.basename(src))
os.makedirs(os.path.dirname(dst), exist_ok=True)
done = set()
if os.path.exists(dst):
done = {(json.loads(l)["id"], json.loads(l)["sample"]) for l in open(dst, encoding="utf-8")}
todo = [r for r in recs if (r["id"], r["sample"]) not in done]
print(src, len(todo), "records to process", flush=True)
with open(dst, "a", encoding="utf-8") as f, ThreadPoolExecutor(args.workers) as ex:
for out in ex.map(run, todo):
f.write(json.dumps(out, ensure_ascii=False) + "\n")
f.flush()
print("done", stats, flush=True)
|