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

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

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"

LFM3-1.7B-inference-modules-experiments / research /scripts /extract_dir_correct.py

HelioAI

LFM3-1.7B inference modules: experimental release

67ebcef verified 2 days ago

Raw History Blame Contribute Delete

3.81 kB

"""'Correct-reasoning' direction from discordant pairs: problems where one of the model's two samples is right and
the other wrong. Difficulty is controlled (same problem), length is controlled (both responses are averaged over the
same number of leading tokens, at most --n-tok). direction[layer] = mean over pairs of (h_correct - h_wrong).
usage: extract_dir_correct.py --gen results/dev_orig/dev.jsonl --graded results/dev_orig/graded.jsonl
                              --items data/train/dev.jsonl --out results/cvec/cor.pt"""
import argparse, json, collections, torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import transformers.integrations.sdpa_attention as _sdpa

# Windows torch has no flash kernel: masked/GQA SDPA would fall back to the O(L^2)-memory math kernel.
# Expanding K/V instead keeps it on the memory-efficient kernel.
_sdpa.use_gqa_in_sdpa = lambda *a, **k: False
ap = argparse.ArgumentParser()
for a in ["--gen", "--graded", "--items", "--out"]:
    ap.add_argument(a, required=True)
ap.add_argument("--model", default="models/LFM3-1.7B")
ap.add_argument("--n-tok", type=int, default=2048)
args = ap.parse_args()
tok = AutoTokenizer.from_pretrained(args.model)
model = AutoModelForCausalLM.from_pretrained(args.model, dtype=torch.bfloat16, attn_implementation="sdpa",
                                             trust_remote_code=True).cuda().eval()
L, H = model.config.num_hidden_layers, model.config.hidden_size
items = {json.loads(l)["id"]: json.loads(l) for l in open(args.items, encoding="utf-8")}
ok = {(g["id"], g["sample"]): g["correct"] for g in map(json.loads, open(args.graded, encoding="utf-8"))}
recs = collections.defaultdict(dict)
for l in open(args.gen, encoding="utf-8"):
    r = json.loads(l)
    if "tokens" in r and (r["id"], r["sample"]) in ok:
        recs[r["id"]][r["sample"]] = r
pairs = []
for pid, s in recs.items():
    if len(s) == 2 and ok[(pid, 0)] != ok[(pid, 1)]:
        c, w = (s[0], s[1]) if ok[(pid, 0)] else (s[1], s[0])
        pairs.append((pid, c, w))
print(len(pairs), "discordant pairs", collections.Counter(items[p]["task"] for p, _, _ in pairs), flush=True)
cap = {}
for k, layer in enumerate(model.model.layers):
    layer.register_forward_hook(lambda m, i, o, k=k: cap.__setitem__(k, (o[0] if isinstance(o, tuple) else o)[0]))


def mean_states(prompt_ids, resp, n):
    ids = prompt_ids + resp[:n]
    with torch.no_grad():
        model(input_ids=torch.tensor([ids], device="cuda"), use_cache=False)
    return torch.stack([cap[k][len(prompt_ids):].float().mean(0) for k in range(L)]).double().cpu()


diff = torch.zeros(L, H, dtype=torch.float64)
norm = torch.zeros(L, dtype=torch.float64)
per_pair = []
for i, (pid, c, w) in enumerate(pairs):
    p = tok.apply_chat_template([{"role": "user", "content": items[pid]["prompt"]}], tokenize=True,
                                add_generation_prompt=True, return_dict=False)
    n = min(len(c["tokens"]), len(w["tokens"]), args.n_tok)
    hc, hw = mean_states(p, c["tokens"], n), mean_states(p, w["tokens"], n)
    diff += hc - hw
    norm += 0.5 * (hc.norm(dim=-1) + hw.norm(dim=-1))
    per_pair.append((hc - hw).float())
    if i % 10 == 0:
        print(i, flush=True)
d = (diff / len(pairs)).float()
avg_norm = (norm / len(pairs)).float()
P = torch.stack(per_pair)  # pairs x L x H
for k in range(L):
    loo = (P[:, k].sum(0, keepdim=True) - P[:, k]) / (len(P) - 1)  # mean of the OTHER pairs (leave-one-out)
    agree = ((P[:, k] * loo).sum(-1) > 0).float().mean()  # held-out agreement; 0.5 = no consistent direction
    print(f"layer {k:2d} |dir| {d[k].norm():6.3f} |h| {avg_norm[k]:6.2f} ratio {d[k].norm() / avg_norm[k]:.3f} pair-agreement {agree:.2f}")
torch.save({"dir": d, "avg_norm": avg_norm, "count": len(pairs)}, args.out)
print("saved", args.out)

	"""'Correct-reasoning' direction from discordant pairs: problems where one of the model's two samples is right and
	the other wrong. Difficulty is controlled (same problem), length is controlled (both responses are averaged over the
	same number of leading tokens, at most --n-tok). direction[layer] = mean over pairs of (h_correct - h_wrong).
	usage: extract_dir_correct.py --gen results/dev_orig/dev.jsonl --graded results/dev_orig/graded.jsonl
	--items data/train/dev.jsonl --out results/cvec/cor.pt"""
	import argparse, json, collections, torch
	from transformers import AutoModelForCausalLM, AutoTokenizer
	import transformers.integrations.sdpa_attention as _sdpa

	# Windows torch has no flash kernel: masked/GQA SDPA would fall back to the O(L^2)-memory math kernel.
	# Expanding K/V instead keeps it on the memory-efficient kernel.
	_sdpa.use_gqa_in_sdpa = lambda a, *k: False
	ap = argparse.ArgumentParser()
	for a in ["--gen", "--graded", "--items", "--out"]:
	ap.add_argument(a, required=True)
	ap.add_argument("--model", default="models/LFM3-1.7B")
	ap.add_argument("--n-tok", type=int, default=2048)
	args = ap.parse_args()
	tok = AutoTokenizer.from_pretrained(args.model)
	model = AutoModelForCausalLM.from_pretrained(args.model, dtype=torch.bfloat16, attn_implementation="sdpa",
	trust_remote_code=True).cuda().eval()
	L, H = model.config.num_hidden_layers, model.config.hidden_size
	items = {json.loads(l)["id"]: json.loads(l) for l in open(args.items, encoding="utf-8")}
	ok = {(g["id"], g["sample"]): g["correct"] for g in map(json.loads, open(args.graded, encoding="utf-8"))}
	recs = collections.defaultdict(dict)
	for l in open(args.gen, encoding="utf-8"):
	r = json.loads(l)
	if "tokens" in r and (r["id"], r["sample"]) in ok:
	recs[r["id"]][r["sample"]] = r
	pairs = []
	for pid, s in recs.items():
	if len(s) == 2 and ok[(pid, 0)] != ok[(pid, 1)]:
	c, w = (s[0], s[1]) if ok[(pid, 0)] else (s[1], s[0])
	pairs.append((pid, c, w))
	print(len(pairs), "discordant pairs", collections.Counter(items[p]["task"] for p, _, _ in pairs), flush=True)
	cap = {}
	for k, layer in enumerate(model.model.layers):
	layer.register_forward_hook(lambda m, i, o, k=k: cap.__setitem__(k, (o[0] if isinstance(o, tuple) else o)[0]))


	def mean_states(prompt_ids, resp, n):
	ids = prompt_ids + resp[:n]
	with torch.no_grad():
	model(input_ids=torch.tensor([ids], device="cuda"), use_cache=False)
	return torch.stack([cap[k][len(prompt_ids):].float().mean(0) for k in range(L)]).double().cpu()


	diff = torch.zeros(L, H, dtype=torch.float64)
	norm = torch.zeros(L, dtype=torch.float64)
	per_pair = []
	for i, (pid, c, w) in enumerate(pairs):
	p = tok.apply_chat_template([{"role": "user", "content": items[pid]["prompt"]}], tokenize=True,
	add_generation_prompt=True, return_dict=False)
	n = min(len(c["tokens"]), len(w["tokens"]), args.n_tok)
	hc, hw = mean_states(p, c["tokens"], n), mean_states(p, w["tokens"], n)
	diff += hc - hw
	norm += 0.5 * (hc.norm(dim=-1) + hw.norm(dim=-1))
	per_pair.append((hc - hw).float())
	if i % 10 == 0:
	print(i, flush=True)
	d = (diff / len(pairs)).float()
	avg_norm = (norm / len(pairs)).float()
	P = torch.stack(per_pair) # pairs x L x H
	for k in range(L):
	loo = (P[:, k].sum(0, keepdim=True) - P[:, k]) / (len(P) - 1) # mean of the OTHER pairs (leave-one-out)
	agree = ((P[:, k] * loo).sum(-1) > 0).float().mean() # held-out agreement; 0.5 = no consistent direction
	print(f"layer {k:2d} \|dir\| {d[k].norm():6.3f} \|h\| {avg_norm[k]:6.2f} ratio {d[k].norm() / avg_norm[k]:.3f} pair-agreement {agree:.2f}")
	torch.save({"dir": d, "avg_norm": avg_norm, "count": len(pairs)}, args.out)
	print("saved", args.out)