LFM2.5-8B-A1B-Wolof-Instruct
Private production artifact — ongoing/experimental and not a public release.
Exact base
LiquidAI/LFM2.5-8B-A1B-Base
Training stage
Instruct. This artifact passed the recorded automated gates in
training_manifest.json; no unrecorded quality claim is made.
Tokenizer family
128k-ext. Tokenizer-family-specific BPB must not be compared as
perplexity across the 65k and 128k families.
Data caveat
Continual pretraining uses the cleaned Wolof corpus protocol. Instruction data is reweighted and excludes the source Hub test split, but benchmark-like examples may have existed in upstream pretraining. Raw private rows are not included here.
Intended use
Private research and evaluation of Wolof language adaptation.
Limitations
This model is experimental. Wolof orthography, code-switching, factuality, reasoning, long-context behavior, and safety have not been comprehensively validated. Native-speaker review remains necessary before broader use.
Evaluation
- Tokenizer family:
128k-ext - Protocol: HF
evaluate_clm.pyBPB is the cross-tokenizer gold metric. - Machine-readable copy:
metrics.json.
Causal LM (verbalized corpus test)
| Metric | Value |
|---|---|
| BPB ↓ | 1.3356 |
| Perplexity | 52.0 |
| BLEU | 3.21 |
| chrF | 19.42 |
Instruct task eval (12_evaluate_sft.py)
| Task | Loss | PPL |
|---|---|---|
| classification | 6.08653345833654 | 439.8938546319482 |
| instruction | 3.2272499946900357 | 25.210233281567458 |
| intent | 4.928517818450928 | 138.1745607052873 |
| ner | 4.99477219581604 | 147.63930870009077 |
| qa | 3.7394813299179077 | 42.076160858575555 |
| tool_use | 2.0756072495822555 | 7.969384395313601 |
| translation | 4.8821809500293005 | 131.91805706725611 |