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:8080/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://127.0.0.1:8080/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://127.0.0.1:8080/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: 15,468 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 | """Procedural curriculum: generators of verifiable reasoning problems with exact answers.
The families are deliberately NOT the families of the eval suites (no BBH task types, no MMLU questions): training on
them and then measuring BBH / MMLU-Pro / math measures transfer of the skill (state tracking, systematic search,
exact computation, careful reading of code), not memorised task formats.
Every item: {id, task: "math", src: "cur/<family>", prompt, answer} with the answer graded by grade.py's math checker.
usage: curriculum.py <n_per_family> <out.jsonl> [seed]
"""
import itertools
import json
import math
import random
import sys
from fractions import Fraction
TAIL = "\n\nPlease reason step by step, and put your final answer within \\boxed{}."
NAMES = ["A", "B", "C", "D", "E", "F", "G", "H", "I"]
def code_trace(r):
"""Short Python program; ask for the printed value (single line)."""
a, b, n = r.randint(2, 9), r.randint(1, 7), r.randint(4, 9)
kind = r.randint(0, 4)
if kind == 0:
src = f"x = {a}\ny = {b}\nfor i in range({n}):\n if i % 2 == 0:\n x = x + y * i\n else:\n y = (y + x) % 11\nprint(x - y)"
elif kind == 1:
lst = [r.randint(-9, 20) for _ in range(n)]
src = f"nums = {lst}\nbest, cur = nums[0], 0\nfor v in nums:\n cur = max(v, cur + v)\n best = max(best, cur)\nprint(best)"
elif kind == 2:
w = "".join(r.choice("abcde") for _ in range(n + 3))
src = f"s = '{w}'\nseen = {{}}\nout = 0\nfor i, ch in enumerate(s):\n if ch in seen:\n out += i - seen[ch]\n seen[ch] = i\nprint(out)"
elif kind == 3:
src = f"def f(k):\n if k <= 1:\n return {b}\n return f(k - 1) + {a} * f(k - 2) % 7\nprint(f({n}))"
else:
lst = [r.randint(1, 30) for _ in range(n)]
src = f"a = {lst}\nfor i in range(len(a)):\n for j in range(len(a) - 1 - i):\n if a[j] % {a} > a[j + 1] % {a}:\n a[j], a[j + 1] = a[j + 1], a[j]\nprint(a[{r.randint(0, n - 1)}])"
env = {}
import io, contextlib
buf = io.StringIO()
with contextlib.redirect_stdout(buf):
exec(src, env) # noqa: S102 (our own generated code)
return f"What does the following Python program print?\n\n```python\n{src}\n```", buf.getvalue().strip()
def linear_system(r):
while True:
x, y, z = (r.randint(-9, 9) for _ in range(3))
rows = [[r.randint(-5, 5) for _ in range(3)] for _ in range(3)]
det = (rows[0][0] * (rows[1][1] * rows[2][2] - rows[1][2] * rows[2][1])
- rows[0][1] * (rows[1][0] * rows[2][2] - rows[1][2] * rows[2][0])
+ rows[0][2] * (rows[1][0] * rows[2][1] - rows[1][1] * rows[2][0]))
if det != 0:
break
eqs = []
for c in rows:
rhs = c[0] * x + c[1] * y + c[2] * z
terms = " ".join(f"{'+' if v >= 0 else '-'} {abs(v)}{s}" for v, s in zip(c, "xyz") if v != 0).lstrip("+ ")
eqs.append(f"{terms} = {rhs}")
p, q = r.randint(1, 5), r.randint(1, 5)
return ("Solve the system of equations\n" + "\n".join(eqs) + f"\nand find the value of {p}x + {q}y - z.",
str(p * x + q * y - z))
def number_theory(r):
k = r.randint(0, 2)
if k == 0:
a, e, m = r.randint(2, 50), r.randint(50, 2000), r.choice([7, 9, 11, 13, 17, 19, 23, 100, 1000])
return f"Find the remainder when ${a}^{{{e}}}$ is divided by ${m}$.", str(pow(a, e, m))
if k == 1:
nums = [r.randint(2, 60) * r.choice([1, 2, 3, 6]) for _ in range(3)]
g = math.gcd(*nums)
l = math.lcm(*nums)
return f"Let $g = \\gcd({nums[0]}, {nums[1]}, {nums[2]})$ and $L = \\operatorname{{lcm}}({nums[0]}, {nums[1]}, {nums[2]})$. Compute $L / g$.", str(l // g)
n = r.randint(200, 5000)
if n < 300:
return f"How many positive divisors does ${n}^2$ have?", str(sum(1 for d in range(1, n * n + 1) if (n * n) % d == 0))
return f"How many positive divisors does ${n}$ have?", str(sum(1 for d in range(1, n + 1) if n % d == 0))
def graph_path(r):
n = r.randint(6, 9)
nodes = NAMES[:n]
edges = {}
for i in range(1, n): # spanning tree keeps it connected
j = r.randint(0, i - 1)
edges[(nodes[j], nodes[i])] = r.randint(1, 12)
for _ in range(r.randint(n, 2 * n)):
a, b = r.sample(nodes, 2)
if (a, b) not in edges and (b, a) not in edges:
edges[(a, b)] = r.randint(1, 12)
dist = {v: math.inf for v in nodes}
dist[nodes[0]] = 0
todo = set(nodes)
while todo:
u = min(todo, key=lambda v: dist[v])
todo.remove(u)
for (a, b), w in edges.items():
for s, t in ((a, b), (b, a)):
if s == u and dist[u] + w < dist[t]:
dist[t] = dist[u] + w
target = max(nodes[1:], key=lambda v: (dist[v], r.random()))
lines = [f"{a} - {b}: {w}" for (a, b), w in r.sample(list(edges.items()), len(edges))]
return (f"Roads connect towns {', '.join(nodes)} (both directions, with lengths):\n" + "\n".join(lines) +
f"\nWhat is the length of the shortest route from {nodes[0]} to {target}?", str(dist[target]))
def scheduling(r):
n = r.randint(5, 8)
tasks = NAMES[:n]
dur = {t: r.randint(1, 9) for t in tasks}
pre = {t: sorted(r.sample(tasks[:i], r.randint(0, min(2, i)))) if i else [] for i, t in enumerate(tasks)}
fin = {}
for t in tasks:
fin[t] = max((fin[p] for p in pre[t]), default=0) + dur[t]
desc = "\n".join(f"Task {t} takes {dur[t]} hours" + (f" and can start only after {' and '.join(pre[t])} "
f"{'is' if len(pre[t]) == 1 else 'are'} finished." if pre[t] else ".")
for t in r.sample(tasks, n))
return ("A project consists of tasks; any number of tasks can run at the same time.\n" + desc +
"\nWhat is the minimum number of hours needed to finish all tasks?", str(max(fin.values())))
def counting(r):
N = r.randint(100, 3000)
a, b = r.sample([2, 3, 4, 5, 6, 7, 9, 11], 2)
c = r.choice([x for x in [3, 4, 5, 7, 8, 10, 13] if x not in (a, b)])
cnt = sum(1 for k in range(1, N + 1) if (k % a == 0 or k % b == 0) and k % c != 0)
return f"How many integers from 1 to {N} (inclusive) are divisible by {a} or by {b}, but not by {c}?", str(cnt)
def probability(r):
k = r.randint(0, 1)
if k == 0:
s = r.randint(5, 10)
faces = r.choice([4, 6, 8])
good = sum(1 for x, y, z in itertools.product(range(1, faces + 1), repeat=3) if x + y + z == s)
p = Fraction(good, faces ** 3)
return f"Three fair {faces}-sided dice (faces 1 to {faces}) are rolled. What is the probability that the sum is {s}? Give the answer as a fraction in lowest terms.", f"\\frac{{{p.numerator}}}{{{p.denominator}}}"
red, blue = r.randint(2, 7), r.randint(2, 7)
d = r.randint(2, 3)
tot = math.comb(red + blue, d)
good = sum(math.comb(red, i) * math.comb(blue, d - i) for i in range(1, d + 1) if i <= red and d - i <= blue)
p = Fraction(good, tot)
return f"A bag has {red} red and {blue} blue marbles. {d} marbles are drawn at random without replacement. What is the probability that at least one of them is red? Give the answer as a fraction in lowest terms.", f"\\frac{{{p.numerator}}}{{{p.denominator}}}"
def cipher(r):
words = ["planet", "garden", "silver", "rocket", "window", "puzzle", "forest", "bridge", "candle", "violin",
"harbor", "meadow", "tunnel", "lantern", "compass"]
w = r.choice(words)
s = r.randint(3, 22)
enc = "".join(chr((ord(ch) - 97 + s) % 26 + 97) for ch in w)
w2 = r.choice([x for x in words if x != w])
enc2 = "".join(chr((ord(ch) - 97 + s) % 26 + 97) for ch in w2)
return (f"A Caesar cipher shifts every letter forward by the same amount (wrapping z to a). The word '{w}' is "
f"encrypted as '{enc}'. Using the same shift, decrypt '{enc2}'. Give the decrypted word.",
f"\\text{{{w2}}}")
def determinant(r):
m = [[r.randint(-6, 6) for _ in range(3)] for _ in range(3)]
d = (m[0][0] * (m[1][1] * m[2][2] - m[1][2] * m[2][1]) - m[0][1] * (m[1][0] * m[2][2] - m[1][2] * m[2][0])
+ m[0][2] * (m[1][0] * m[2][1] - m[1][1] * m[2][0]))
rows = " \\\\ ".join(" & ".join(str(v) for v in row) for row in m)
return f"Compute the determinant of the matrix $\\begin{{pmatrix}} {rows} \\end{{pmatrix}}$.", str(d)
def rates(r):
a, b = r.randint(2, 12), r.randint(2, 12)
c = r.randint(a * b // (a + b) + 2, 40) if r.random() < 0.5 else None
if c:
t = Fraction(1, Fraction(1, a) + Fraction(1, b) - Fraction(1, c))
q = f"Pipe 1 fills a tank in {a} hours and pipe 2 fills it in {b} hours, while a drain empties the full tank in {c} hours. Starting empty with both pipes and the drain open, how many hours until the tank is full?"
else:
t = Fraction(1, Fraction(1, a) + Fraction(1, b))
q = f"Alice paints a fence in {a} hours and Bob paints the same fence in {b} hours. Working together, how many hours do they need?"
return q + " Give an exact answer.", (str(t.numerator) if t.denominator == 1 else f"\\frac{{{t.numerator}}}{{{t.denominator}}}")
def recurrence(r):
a, b, c = r.randint(1, 4), r.randint(-3, 3), r.randint(-5, 5)
x0, x1, n = r.randint(0, 5), r.randint(0, 5), r.randint(8, 14)
xs = [x0, x1]
for _ in range(n - 1):
xs.append(a * xs[-1] + b * xs[-2] + c)
return (f"A sequence is defined by $x_0 = {x0}$, $x_1 = {x1}$ and $x_{{n+1}} = {a}x_n {'+' if b >= 0 else '-'} "
f"{abs(b)}x_{{n-1}} {'+' if c >= 0 else '-'} {abs(c)}$ for $n \\ge 1$. Find $x_{{{n}}}$.", str(xs[n]))
def string_ops(r):
"""Hard: state tracking over a sequence of string operations."""
s = "".join(r.choice("abcdefghij") for _ in range(r.randint(7, 10)))
ops, cur = [], list(s)
for _ in range(r.randint(7, 11)):
k = r.randint(0, 4)
if k == 0:
i, j = sorted(r.sample(range(len(cur)), 2))
ops.append(f"swap the letters at positions {i + 1} and {j + 1}")
cur[i], cur[j] = cur[j], cur[i]
elif k == 1:
t = r.randint(1, len(cur) - 1)
ops.append(f"rotate the string left by {t}")
cur = cur[t:] + cur[:t]
elif k == 2:
i, j = sorted(r.sample(range(len(cur)), 2))
ops.append(f"reverse the part from position {i + 1} to position {j + 1}")
cur[i:j + 1] = cur[i:j + 1][::-1]
elif k == 3:
a, b = r.sample("abcdefghij", 2)
ops.append(f"replace every '{a}' with '{b}'")
cur = [b if c == a else c for c in cur]
else:
i = r.randint(0, len(cur) - 1)
ops.append(f"delete the letter at position {i + 1}")
del cur[i]
steps = "\n".join(f"{n + 1}. {o}" for n, o in enumerate(ops))
return (f"Start with the string '{s}' (positions are counted from 1). Apply these operations in order:\n{steps}\n"
"What is the final string?", f"\\text{{{''.join(cur)}}}")
def stack_sim(r):
st, ops, total = [], [], 0
for _ in range(r.randint(12, 18)):
if st and r.random() < 0.4:
k = r.choice(["pop", "dup", "add"] if len(st) >= 2 else ["pop", "dup"])
if k == "pop":
total += st.pop()
ops.append("pop (add the popped value to the running total)")
elif k == "dup":
st.append(st[-1])
ops.append("duplicate the top")
else:
a, b = st.pop(), st.pop()
st.append(a + b)
ops.append("pop two values and push their sum")
else:
v = r.randint(1, 20)
st.append(v)
ops.append(f"push {v}")
steps = "\n".join(f"{n + 1}. {o}" for n, o in enumerate(ops))
return (f"A stack starts empty and a running total starts at 0. Perform:\n{steps}\nAt the end, what is the running "
f"total plus the sum of all values left on the stack times 2?", str(total + 2 * sum(st)))
def crt(r):
ms = r.sample([3, 4, 5, 7, 9, 11, 13], 3)
while math.gcd(ms[0], ms[1]) != 1 or math.gcd(ms[0], ms[2]) != 1 or math.gcd(ms[1], ms[2]) != 1:
ms = r.sample([3, 4, 5, 7, 9, 11, 13], 3)
x = r.randint(0, ms[0] * ms[1] * ms[2] - 1)
cond = ", ".join(f"$x \\equiv {x % m} \\pmod{{{m}}}$" for m in ms)
return f"Find the smallest non-negative integer $x$ such that {cond}.", str(x)
def graph_path_big(r):
q, a = graph_path(_Big(r))
return q, a
class _Big:
"""Random proxy that makes graph_path draw 10-12 nodes."""
def __init__(self, r):
self.r = r
self.first = True
def randint(self, a, b):
if self.first and (a, b) == (6, 9):
self.first = False
return self.r.randint(9, 9)
return self.r.randint(a, b)
def __getattr__(self, k):
return getattr(self.r, k)
def counting_big(r):
N = r.randint(10 ** 5, 10 ** 6)
a, b, c = r.sample([2, 3, 5, 7, 11, 13], 3)
cnt = N // a + N // b + N // c - N // (a * b) - N // (a * c) - N // (b * c) + N // (a * b * c)
return f"How many integers from 1 to {N} are divisible by at least one of {a}, {b}, {c}?", str(cnt)
def linear4(r):
while True:
sol = [r.randint(-6, 6) for _ in range(4)]
M = [[r.randint(-4, 4) for _ in range(4)] for _ in range(4)]
A = [row[:] for row in M]
det = Fraction(1)
for i in range(4): # Gaussian elimination determinant
p = next((k for k in range(i, 4) if A[k][i] != 0), None)
if p is None:
det = 0
break
if p != i:
A[i], A[p] = A[p], A[i]
det = -det
det *= A[i][i]
for k in range(i + 1, 4):
f = Fraction(A[k][i], A[i][i])
A[k] = [A[k][j] - f * A[i][j] for j in range(4)]
if det != 0:
break
eqs = []
for row in M:
rhs = sum(c * v for c, v in zip(row, sol))
terms = " ".join(f"{'+' if c >= 0 else '-'} {abs(c)}{s}" for c, s in zip(row, "wxyz") if c).lstrip("+ ")
eqs.append(f"{terms} = {rhs}")
return "Solve the system\n" + "\n".join(eqs) + "\nand give the value of w + 2x + 3y + 4z.", \
str(sol[0] + 2 * sol[1] + 3 * sol[2] + 4 * sol[3])
FAMILIES = [code_trace, linear_system, number_theory, graph_path, scheduling, counting, probability, cipher,
determinant, rates, recurrence]
HARD = [string_ops, stack_sim, crt, graph_path_big, counting_big, linear4]
if __name__ == "__main__":
n, out = int(sys.argv[1]), sys.argv[2]
rng = random.Random(int(sys.argv[3]) if len(sys.argv) > 3 else 0)
fams = HARD if len(sys.argv) > 4 and sys.argv[4] == "hard" else FAMILIES
rows = []
for fam in fams:
seen = set()
while sum(1 for x in rows if x["src"] == "cur/" + fam.__name__) < n:
q, a = fam(rng)
if q in seen:
continue
seen.add(q)
rows.append({"id": f"cur/{fam.__name__}/{len(seen)}", "task": "math", "src": "cur/" + fam.__name__,
"prompt": q + TAIL, "answer": a})
rng.shuffle(rows)
with open(out, "w", encoding="utf-8") as f:
for x in rows:
f.write(json.dumps(x, ensure_ascii=False) + "\n")
print(len(rows), "items")
|