---

library_name: gguf

pipeline_tag: text-generation

tags:

• gguf
• llama.cpp
• wolof
• lfm2.5
• private-production-artifact

base_model: Tonic/LFM2.5-2.6B-Wolof-CPT-v3

model-index:

• name: LFM2.5-2.6B-Wolof-CPT-v3-GGUF

results:

- task:

type: text-generation

name: Wolof GGUF Q4_K_M

dataset:

type: wolof-clm-corpus-test

name: Wolof CLM corpus test (verbalized, gguf-llama-cpp)

split: test

metrics:

- type: perplexity

value: 17.6602

name: llama.cpp perplexity (Q4_K_M)

- type: bleu

value: 3.11

name: BLEU via llama-completion

- type: chrf

value: 16.47

name: chrF via llama-completion

source:

name: GalsenAI LFM2.5 Wolof family eval

url: https://huggingface.co/Tonic/LFM2.5-2.6B-Wolof-CPT-v3-GGUF/blob/main/metrics.json

language:

• wo

---

LFM2.5-2.6B-Wolof-CPT-v3-GGUF

Private production artifact (ongoing/experimental).

• Exact base: Tonic/LFM2.5-2.6B-Wolof-CPT-v3
• Training stage: CPT (+ GGUF export)
• Tokenizer family: see base model card
• Data caveat: Wolof CPT/SFT mixture; no public leakage claims
• Limitations: experimental; verify generation quality before downstream use
• Private production artifact for gated llama.cpp deployment

Files

GGUF quant matrix following LiquidAI parity: BF16, F16, Q8_0, Q6_K, Q5_K_M, Q4_K_M, Q4_0.

Run

llama-cli -hf <namespace>/LFM2.5-2.6B-Wolof-CPT-v3-GGUF:Q4_K_M -p "Salaam aleekum"

<!-- FAMILY_EVAL_START -->

Evaluation

• Tokenizer family: 128k-ext
• Protocol: HF evaluate_clm.py BPB is the cross-tokenizer gold metric.
• Machine-readable copy: metrics.json.

GGUF Q4_K_M (gguf-llama-cpp; not HF-BPB-equivalent)

| Metric | Value |

|---|---:|

| Perplexity | 17.6602 |

| BLEU | 3.11 |

| chrF | 16.47 |

Compare GGUF PPL only within the same GGUF/tokenizer family. Use HF BPB for cross-model ranking.

HF reference (same verbalized corpus)

| Metric | Value |

|---|---:|

| BPB | 1.3295 |

| BLEU | 4.57 |

| chrF | 22.87 |

<!-- FAMILY_EVAL_END -->