LFM2.5 Wolof Instruct-v3

GalsenAI · LiquidAI LFM2.5 · five sizes · ZeroGPU

GalsenAI · LFM2.5 Wolof Instruct-v3

Five LiquidAI LFM2.5 checkpoints adapted to Wolof by GalsenAI and published under Tonic/LFM2.5-*-Wolof-Instruct-v3:

  1. Stock Base → tokenizer extension (65k-ext or 128k-ext)
  2. CPT-v3 — continual pretraining on the ortho non-standard + mixed Wolof pack
  3. Instruct-v3 — SFT on the reweighted Wolof instruction mixture
  4. Eval — HF evaluate_clm.py BPB (bits/byte, lower better) is gold

How to prompt

All five are chat models. This Space uses each repo’s chat_template.jinja via tokenizer.apply_chat_template (<|im_start|> / <|im_end|>).

How to read scores

  • BPB is comparable across tokenizer families; PPL is not.
  • Deltas vs stock Base are shown on each card (from FAMILY_EVAL.md).
  • Native-speaker review still required before product use.

Set Space secret HF_TOKEN with read access to private Tonic/LFM2.5-*-Wolof-Instruct-v3 repos.
Model

All five are Instruct-v3 chat checkpoints. 8B-A1B uses ZeroGPU xlarge.

Defaults: temp=0.7, top-p=0.9, top-k 50, mild rep=1.05. Set top-k=0 to disable top-k.
16 512
0 1.5
0.1 1
0 200
1 1.5

Select a model and generate.

1.2B · Instruct-v3 (recommended)

Field Value
Hub Tonic/LFM2.5-1.2B-Wolof-Instruct-v3
Base CPT Tonic/LFM2.5-1.2B-Wolof-CPT-v3
Size ~1.2B · arch=lfm2 · tok=65k-ext
ZeroGPU large
BPB (↓ better) 1.4263 (-1.645 vs stock 3.071)
BLEU 3.4
chrF 16.52
PPL 11.41 (same-tokenizer only)

Hero size: CPT-v3 (ortho non-standard+mixed pack) → SFT Instruct-v3. Best balance of quality, latency, and ZeroGPU fit.

Strong Wolof instruct checkpoint. BPB 1.43 vs stock base 3.07 (~−1.65). Prefer this for demos unless you need the 2.6B BPB win or tiny on-device sizes.

Prompting: apply_chat_template with each repo’s Jinja template (<|im_start|> roles). System is optional.
Examples (all five Instruct-v3 sizes)
Model System User message Max new tokens Temperature Top-p Top-k (0 = off) Repetition penalty