Fabliq-8B-Agent-FromBase 🌊🔬
Ablation variant of Fabliq-8B-Agent — fine-tuned directly from raw
LiquidAI/LFM2.5-8B-A1B, skipping the ToolBench foundation. Used to isolate the effect of the ToolBench intermediate stage.
🔬 Why this variant?
Fabliq-8B-Agent is trained as: LiquidAI/LFM2.5-8B-A1B → ToolBench-Full-SFT-1Epoch → Fable-5.
This model skips the middle step: LiquidAI/LFM2.5-8B-A1B → Fable-5 (direct).
Comparing the two answers the question: does ToolBench foundation actually help, or does Fable-5 alone give you the same agent?
🧪 Model details
| Architecture | Lfm2MoeForCausalLM (24 layers, 32 experts, 4 experts/token) |
| Parameters | ~8B total / ~1B active (MoE) |
| Context | 8,192 trained · 128K native |
| Precision | bfloat16 |
| Fine-tune type | Full-parameter SFT (direct from base, no ToolBench) |
| License | Apache 2.0 |
📚 Training data
Same as Phase-1: Fable-5-traces, 4,047 rows × 3 epoch.
🔧 Training procedure
| Hyperparameter | Value |
|---|---|
| Schedule | 3 epochs, constant LR |
| Max sequence length | 8,192 |
| Per-device batch size | 2 |
| Gradient accumulation | 4 |
| GPUs | 8× H200 (effective batch 64) |
| Learning rate | 1e-6 (higher than Phase-1's 5e-7, since starting from base) |
| Precision | bf16 |
| Final train_loss | logged in run_config.json |
| Train runtime | ~14 min |
🌳 Model tree
LiquidAI/LFM2.5-8B-A1B
├─ LLM-OS-Models/LFM2.5-8B-A1B-Terminal-ToolBench-Full-SFT-1Epoch
│ └─ LLM-OS-Models/Fabliq-8B-Agent ← Phase-1 (ToolBench → Fable-5)
└─ LLM-OS-Models/Fabliq-8B-Agent-FromBase ← this model (base → Fable-5, direct)
└─ LLM-OS-Models/Fabliq-8B-Agent-FromBase-Reasoning ← + reasoning expansion
🚀 Usage
See Fabliq-8B-Agent — same system prompt, same tool-call format, same inference code.
📜 License
Apache 2.0.