Exp 5C)Fable (fable-mistral-small-4-119b-telco-lora) is a high-rank (r = 32, alpha = 64) BF16 LoRA adapter trained on mistralai/Mistral-Small-4-119B-2603 (119B total / 6.5B active Mixture-of-Experts with Multi-head Latent Attention) for 3GPP Release 17/18 5G Standalone (SA) Core Architecture, Cross-Layer Root-Cause Analysis (RAN / Transport / Core), and Sovereign Telecom Cloud Operations.
It is part of the Mona, Lisa & Fable Trilogy (Exp 5A–5C):
AIHeikki/mona-qwen3.5-4b-epistemic-sparring (Exp 5A): Textual & RAG Epistemic Discipline (Qwen3.5-4B).AIHeikki/lisa-gemma4-e4b-visual-auditor (Exp 5B): Multimodal Slide & Visual Brand Forensics (Gemma-4-E4B-it).AIHeikki/fable-mistral-small-4-119b-telco-lora (Exp 5C, this repository): Frontier-scale (119B MoE) 5G SA Telco Architect & NOC Root-Cause Commander.0% External Frontier LLM Distillation): Every training trajectory in the 240-scenario curriculum was generated by a deterministic Python 3GPP Release 17/18 physics, link-budget, Erlang-C, and state-machine solver (build_telco_curriculum.py) and formatted by mistralai/Mistral-Small-4-119B-2603 itself under reasoning_effort="none". No OpenAI, Anthropic, or Google frontier model outputs were used anywhere in the training corpus (see CLEAN_LINEAGE_ATTESTATION.json).peft BF16 LoRA (r = 32, lora_alpha = 64, 219.8 MB .safetensors) applied across all 288 Multi-head Latent Attention (q_a_proj, q_b_proj, kv_a_proj_with_mqa, kv_b_proj, o_proj) and Shared-Expert MLP (gate_proj, up_proj, down_proj) projections across all 36 layers on an 8x NVIDIA H100 80GB HBM3 node (640 GB total VRAM).1.2810 → 0.0365 (-97.15%, token accuracy 67.9% → 98.98%) in 26.2 minutes on 8x H100.30 Unseen 5G SA Scenarios, Seed 20260405)Evaluated on 30 unseen 5G SA scenarios across all 6 telecom domains (T1 3GPP Rel-17/18 Signaling & Cause Codes, T2 Cross-Layer Fault Localization, T3 RF Link Budget & PRACH Math, T4 UPF/eBPF/SR-IOV Dimensioning, T5 Sovereign Cloud Blast-Radius & Chaos Engineering, T6 Epistemic Telemetry Abstention) against both Base Mistral-Small-4-119B-2603 and Base + 6,500-token Telecom Playbook Prompt (RAG+Prompt):
Held-Out Metric (N = 30 Unseen Scenarios) |
Arm 1: Base Mistral-Small-4-119B |
Arm 2: Base + 6.5k-Token Telco Playbook (RAG+Prompt) |
Arm 3: Fable 119B Telco LoRA (Exp 5C) |
Delta vs Base | Delta vs RAG+Prompt |
|---|---|---|---|---|---|
| Overall Deterministic Rubric Score | 34.94% |
59.94% |
85.22% |
+50.28 pp |
+25.28 pp |
Exact Numerical Engineering Math Tolerance (±1%) |
41.67% |
56.67% |
72.50% |
+30.83 pp |
+15.83 pp |
| 3GPP IE / Timer / Cause-Code Recall | 42.22% |
46.67% |
82.22% |
+40.00 pp |
+35.55 pp |
| Epistemic Abstention & Missing-PCAP Gate | 16.67% |
66.67% |
100.00% |
+83.33 pp |
+33.33 pp |
| Mandatory 5-Block NOC Architectural Compliance | 0.00% |
93.33% |
100.00% |
+100.00 pp |
+6.67 pp |
Blind 3-Way Judge Overall Score (/ 10) |
5.84 / 10 |
7.12 / 10 |
9.52 / 10 |
+3.68 |
+2.40 |
Blind Pairwise Win / Tie / Loss (Fable vs Opponent) |
29 W / 0 T / 1 L (96.7%) |
28 W / 1 T / 1 L (93.3%) |
— | — | — |
.
├── README.md # This Model Card
├── adapter_model.safetensors # 219.8 MB BF16 LoRA weights (r=32, alpha=64 across 288 MLA + Shared-Expert modules)
├── adapter_config.json # PEFT LoRA configuration
├── CLEAN_LINEAGE_ATTESTATION.json # Cryptographic & provenance attestation (0% external LLM distillation)
├── config.json # Mistral-Small-4-119B-2603 base architecture config
├── generation_config.json # Deterministic generation parameters
├── processor_config.json # Processor configuration
├── tekken.json # Mistral Tekken tokenizer vocabulary
├── tokenizer.json # Hugging Face fast tokenizer
├── tokenizer_config.json # Tokenizer config
├── chat_template.jinja # Chat template supporting reasoning_effort ("none" | "high")
└── reports/
├── FABLE_MISTRAL4_119B_TELCO_REPORT.md # Full engineering & 3-arm benchmark report
├── deterministic_3arm_summary.json # Per-scenario deterministic solver scores across all 30 holdout tasks
└── blind_3way_telco_report.json # Blind 3-way judge rankings & rationales
transformers + peft)import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
BASE_ID = "mistralai/Mistral-Small-4-119B-2603"
LORA_ID = "AIHeikki/fable-mistral-small-4-119b-telco-lora"
tokenizer = AutoTokenizer.from_pretrained(LORA_ID, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
BASE_ID,
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True,
)
model = PeftModel.from_pretrained(base_model, LORA_ID)
model.eval()
Released under the Apache 2.0 License, matching the upstream mistralai/Mistral-Small-4-119B-2603 license.
mistralai/Mistral-Small-4-119B-2603