GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
mohamedlotfy50
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
mohamedlotfy50
Model Tree
Qwen/Qwen3.5-0.8B
Input Modalities
Output Modalities
Supported Functionality
One fixed rendering for every benchmark (State / Question / Options with explicit keys,
"Reply with exactly one option key"). Every option is scored as the length-normalized
log-probability of its key tokens (summed key-token logprob ÷ key token count,
acc_norm convention), softmaxed across options. No truncation, no option filtering, no
prompt tuning, no per-benchmark hacks; the policy is declared in environment.json and
in the engine's provenance.
| Area | skill |
|---|---|
| Retrieval & Classification | 0.455 |
| Tools & Automation | 0.360 |
| Knowledge & Reasoning | 0.185 |
| Language Understanding | 0.226 |
| Arts & Human Taste | 0.060 |
Selected families (raw): PhishNChips 0.999 · ARC-Easy 0.859 · FinEntity 0.857 · CLINC150 0.728 ★ · BFCL 0.796 ★ · BANKING77 0.680 ★ · GSM8K 0.717 · HellaSwag 0.730 ★ · ARC-Challenge 0.688 · ToolRet 0.506 · API-Bank 0.408 ★ · SGD 0.604.
(★ = gold benchmarks, ×1.2 weight.)
python
import json, math, torchfrom transformers import AutoProcessor, AutoModelForMultimodalLM model_id = "mohamedlotfy50/sifr-0.8b-v3"processor = AutoProcessor.from_pretrained(model_id)tok = processor.tokenizermodel = AutoModelForMultimodalLM.from_pretrained( model_id, dtype=torch.bfloat16, device_map="cuda:0").eval() SYSTEM = ("You are a decision engine. Read the state, then answer the question " "by replying with exactly one option key from the list. Do not explain.") def render(state, q): s = state if isinstance(state, str) else json.dumps( state, ensure_ascii=False, separators=(",", ":")) body = ["State:", s if s else "(empty)", "", "Question:", q["instructions"], "", "Options:"] for key, desc in q["criteria"].items(): body.append(f"{key}: {key if desc is None else desc}") body += ["", "Reply with exactly one option key."] prompt = tok.apply_chat_template( [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "\n".join(body)}], tokenize=False, add_generation_prompt=True) return tok(prompt, add_special_tokens=False)["input_ids"] @torch.inference_mode()def score(prompt_ids, criteria): """Length-normalized P(key | prompt) for every option.""" keys, sums, lens = list(criteria), [], [] for k in keys: kids = tok(k, add_special_tokens=False)["input_ids"] out = model(input_ids=torch.tensor([prompt_ids + kids], device=model.device), logits_to_keep=len(kids) + 1) lsm = torch.log_softmax(out.logits[0].float(), dim=-1) sums.append(sum(lsm[t, kt].item() for t, kt in enumerate(kids))) lens.append(len(kids)) norm = [s / l for s, l in zip(sums, lens)] top = max(norm); w = [math.exp(x - top) for x in norm]; z = sum(w) return {k: wi / z for k, wi in zip(keys, w)} state = {"task": "sentiment classification"}q = {"instructions": "Classify the sentiment of this review:\n" "One star. The screen arrived cracked and support never replied.", "criteria": {"A": "negative", "B": "positive"}}probs = score(render(state, q), q["criteria"])# -> {'A': 0.755, 'B': 0.245}print("choice:", max(probs, key=probs.get)) # -> ANotes:
score() convention is length-normalized (summed key-token logprob ÷ token
count); for equal-length option sets it reduces to plain summed logprobs.{"false": "False", "true": "True"}.sifr_engine.py in this repo plugs in directly as --engine sifr_engine:SifrEngine.Typed decisions with calibrated probabilities where the answer space is explicit: classification, retrieval verdicts, tool selection, verification (supported/not), routing, binary judgment. The engine returns the full option distribution, so thresholds, abstention and cascades can be built on top.
model.safetensors — the checkpoint (best_7800)sifr_engine.py — the Decision-Index engine (policy in provenance)probe.json, temperatures.json, tokenizer.json, config.json — supporting filesbibtex
@misc{sifr2026, title={Sifr: a calibrated System-1 decision engine at 0.8B}, author={Lotfy, Mohamed}, year={2026}, howpublished={\url{https://huggingface.co/mohamedlotfy50/sifr-0.8b-v3}}}License: Apache 2.0 for our additions; the base model's license applies to the inherited weights and tokenizer.
import json, math, torchfrom transformers import AutoProcessor, AutoModelForMultimodalLM model_id = "mohamedlotfy50/sifr-0.8b-v3"processor = AutoProcessor.from_pretrained(model_id)tok = processor.tokenizermodel = AutoModelForMultimodalLM.from_pretrained( model_id, dtype=torch.bfloat16, device_map="cuda:0").eval() SYSTEM = ("You are a decision engine. Read the state, then answer the question " "by replying with exactly one option key from the list. Do not explain.") def render(state, q): s = state if isinstance(state, str) else json.dumps( state, ensure_ascii=False, separators=(",", ":")) body = ["State:", s if s else "(empty)", "", "Question:", q["instructions"], "", "Options:"] for key, desc in q["criteria"].items(): body.append(f"{key}: {key if desc is None else desc}") body += ["", "Reply with exactly one option key."] prompt = tok.apply_chat_template( [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "\n".join(body)}], tokenize=False, add_generation_prompt=True) return tok(prompt, add_special_tokens=False)["input_ids"] @torch.inference_mode()def score(prompt_ids, criteria): """Length-normalized P(key | prompt) for every option.""" keys, sums, lens = list(criteria), [], [] for k in keys: kids = tok(k, add_special_tokens=False)["input_ids"] out = model(input_ids=torch.tensor([prompt_ids + kids], device=model.device), logits_to_keep=len(kids) + 1) lsm = torch.log_softmax(out.logits[0].float(), dim=-1) sums.append(sum(lsm[t, kt].item() for t, kt in enumerate(kids))) lens.append(len(kids)) norm = [s / l for s, l in zip(sums, lens)] top = max(norm); w = [math.exp(x - top) for x in norm]; z = sum(w) return {k: wi / z for k, wi in zip(keys, w)} state = {"task": "sentiment classification"}q = {"instructions": "Classify the sentiment of this review:\n" "One star. The screen arrived cracked and support never replied.", "criteria": {"A": "negative", "B": "positive"}}probs = score(render(state, q), q["criteria"])# -> {'A': 0.755, 'B': 0.245}print("choice:", max(probs, key=probs.get)) # -> A@misc{sifr2026, title={Sifr: a calibrated System-1 decision engine at 0.8B}, author={Lotfy, Mohamed}, year={2026}, howpublished={\url{https://huggingface.co/mohamedlotfy50/sifr-0.8b-v3}}}