LFM-2.6B-Claude4.8-GPT-5.6-Sol-Grok-4.6_Fable5-Distilled
OpenIntelligenceNet/LFM-2.6B-Claude4.8-GPT-5.6-Sol-Grok-4.6_Fable5-Distilled is a 2.6-billion parameter instruction-tuned edge model. It is built upon huihui-ai/Huihui-LFM2-2.6B-Exp-abliterated and fine-tuned on a 200,000-sample multi-domain distillation corpus.
Model Overview
This model combines edge-device inference speeds with high-grade synthetic reasoning distilled from leading frontier models (Claude 4.8, GPT-5.6, Sol, Grok 4.6, and Fable-5). Refusal patterns and boilerplate safety artifacts were purged from the training set, allowing the model to answer direct technical questions without evasive disclaimers.
- Base Model: huihui-ai/Huihui-LFM2-2.6B-Exp-abliterated* Parameters: 2.6B
- Context Length: 2,048 tokens
- Format: ChatML (
<|im_start|user / assistant<|im_end|>) - Precision: FP16 Merged Standalone Weights
Dataset Breakdown & Distribution
The model was trained on 200,000 deduplicated, format-verified conversational pairs across 6 core technical domains:
| Domain | Target Samples | Share | Sources & Focus Areas |
|---|---|---|---|
| Coding & Systems | 60,000 | 30.0% | DeepSeek, Fable-5, Genesis, Kimi, algorithmic code generation |
| Distilled Frontier Reasoning | 50,000 | 25.0% | Chain-of-thought distillation from GPT-5.6, Grok 4.6, Claude 4.8, Mythos |
| Cybersecurity | 25,000 | 12.5% | Exploit analysis, vulnerability research, penetration testing mechanics |
| General Instruction | 25,000 | 12.5% | Multi-turn instruction following, structured formatting, Alpaca / Dolly |
| Science & Medicine | 24,000 | 12.0% | Biology, medical rationale, physics, chemistry, CS theory |
| Mathematics | 16,000 | 8.0% | MetaMathQA step-by-step arithmetic and algebraic proofs |
Training Methodology
- Hardware: 2x NVIDIA Tesla T4 GPUs via PyTorch DDP (
torchrun) - Optimization: LoRA Rank r=64, Alpha=64, targeting all attention & MLP projection layers
- Optimizer: 8-bit AdamW with Cosine Learning Rate Schedule (3e-5 peak LR)
- Packing: 1,024-token dense 1D sequence packing (zero pad-token compute waste)
- Loss Trajectory: Converged from initial loss down to ~0.76 across the full dataset
Inference with Transformers
``lpython import torch from transformers import AutoModelForCausaLLM, AutoTokenizer
model_id = "OpenIntelligenceNet/LFM-2.6B-Claude4.8-GPT-5.6-Sol-Grok-4.6_Fable5-Distilled"
tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.float16, device_map="auto" )
messages = [ {"role": "user", "content": "Explain how buffer overflow vulnerabilities occur at the memory level and how to prevent them in C."} ]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=512, temperature=0.7, top_p=0.9, repetition_penalty=1.15, eos_tokenid=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id )
new_tokens = outputs[0][inputs.input_ids.shape[1]:] response = tokenizer.decode(new_tokens, skip_special_tokens=True) print(response)