AfriSLM-2B Pruned & Context-Calibrated (Kronos q2c / nf3g128)

Pruned, Hadamard-GPTQ quantized, and context-calibrated builds of qvac/TranslatePsy-AfriSLM-2B (Apache-2.0 fine-tune of Qwen/Qwen3.5-2B) compiled in 100% pure Rust for on-device translation in Kronos and the Translate keyboard (iOS & Android).

What Changed vs. Base Checkpoint (5.19 GiB -> 572โ€“671 MiB .lzfse)

  1. Structural & BPE Vocabulary Pruning (prune_vocab, 248,320 -> 179,008 tokens):
    • Stripped all 297 unused multimodal vision (visual.*) and Multi-Token Prediction (mtp.*) tensors.
    • Pruned source-code-only BPE merges and ultra-rare high-rank compound tokens while retaining 100% of African tonal/subdot/hooked orthographies (แบน, แป, แนฃ, แปฅ, แป‹, ษ“, ษ—, ฦ™, ฦด, ษ›, ษ”, combining tones, โ‚ฆ), East Asian (Chinese Hanzi, Japanese Kanji/Kana, Korean Hangul), Arabic/Persian/Urdu, Devanagari/Bengali, Cyrillic, Greek, Thai, Ethiopic, European/Vietnamese diacritics, and emojis, closed under BPE merge ancestors and padded to a multiple of 64.
    • Deduplicated tied input/output embeddings (CACHE_MAGIC_V2 K2TQW35B2, embed = None sharing unbiased 4-bit Hadamard RTN NF4 lm_head), saving ~205 MiB with zero accuracy loss.
  2. Differentiable Context + BF16 Teacher-KL Calibration (train_qwen35_context):
    • Trained Layers 19..24 (NfSdotMatrix per-group scales, DeltaNet recurrent gates/norms, QKNorm) + final_norm in pure Rust against conversational [Background Information] examples (pronoun/addressee/object gender, formal/casual register, polysemy, target-language context, Nigerian Pidgin & Nigerian/African pairs) + exact top-64 KL distillation from the unquantized BF16 teacher across 36 languages, written back in-place into BF16 scales with 0 extra bytes.

Model Size Tiers Included

Both .k2tb containers use the self-describing K2TQW35B2 format and support zero-copy mmap execution (kronos_set_zero_copy_weights(1)) on iOS and Android:

Variant Precision Raw .k2tb .k2tb.lzfse Download tokenizer.json.lzfse Total Download Compression vs 5.19 GiB
kronos-q2c.k2tb 2-bit Gosset E8Q2 MLP + 3-bit NF3 Mixer + 4-bit NF4 tied head 585.8 MiB (614.2 MB) 571.5 MiB (599.3 MB) 2.4 MiB (2.55 MB) 574.0 MiB 8.9x smaller
kronos-nf3g128.k2tb 3-bit NF3 GPTQ Body + 4-bit NF4 tied head (3.35 bpw) 693.8 MiB (727.5 MB) 671.5 MiB (704.1 MB) 2.4 MiB (2.55 MB) 673.9 MiB 7.5x smaller

Translation Accuracy

Context-Aware Disambiguation (context_ab & 361-Case Multilingual context_eval)

  • context_ab (19 conversational gender/formality/sense/leakage cases):
    • Uncalibrated GPTQ baseline: 3/15 context used (20.0%).
    • Calibrated q2c & nf3g128: 10/15 context used (66.7%, +233% relative) ยท 19/19 (100%) leak-free (identical on Metal GPU and ARM64 NEON --cpu).
  • Held-Out 361-Case Multilingual context_eval (nf3g128, 32 languages):
    • Overall accuracy jumped from 34.2% without context -> 40.0% with context (+5.8%), with 66/66 (100%) no_leak pass rate.
    • Per-language context gains: Chinese +36% (36% -> 73%), Japanese +11% (44% -> 56%), French +24% (29% -> 52%), Spanish +19% (33% -> 52%), Italian +18%, Arabic +17%, Turkish +29%, Dutch +57%, Formal register +29% (29% -> 58%), Object gender +13% (67% -> 80%).

Held-Out FLORES-200 (devtest, English -> Target, chrF++ / BLEU)

Code Language v1 Hy-MT2-1.8B (q2c) v2 AfriSLM-2B (q2c, 574 MiB) v2 AfriSLM-2B (nf3g128, 674 MiB)
HA Hausa 8.7 / 1.8 42.9 / 16.7 51.5 / 25.7
ZU Zulu 12.2 / 2.0 41.1 / 13.2 48.6 / 20.0
IG Igbo 14.1 / 2.3 34.1 / 12.7 41.0 / 17.3
YO Yoruba 9.0 / 1.6 26.4 / 8.5 30.2 / 11.6
SW Swahili 34.0 / 9.6 48.6 / 21.1 56.8 / 29.1
AM Amharic 1.0 / 0.5 8.4 / 1.9 25.3 / 5.9
ZH Chinese (Simplified) 42.3 / 42.4 32.2 / 31.7 37.7 / 37.4
ZT Chinese (Traditional) 21.2 / 17.5 25.3 / 21.4 29.4 / 28.1
JA Japanese 45.0 / 46.0 24.9 / 22.4 30.2 / 28.5
FR French 65.0 / 42.3 49.5 / 26.7 56.1 / 30.7
PT Portuguese 66.8 / 45.2 51.0 / 26.9 55.2 / 30.9
SV Swedish 57.0 / 30.8 โ€” 50.7 / 24.7
RU Russian 52.0 / 26.6 39.3 / 15.8 49.1 / 23.8
VI Vietnamese 55.7 / 35.4 43.0 / 22.5 48.6 / 29.6
RO Romanian 53.9 / 28.3 โ€” 47.3 / 21.0
DE German 59.5 / 34.2 42.5 / 16.1 47.1 / 20.7
ES Spanish 53.7 / 27.8 44.5 / 17.0 45.1 / 17.4
AR Arabic 49.0 / 21.0 35.3 / 11.4 44.4 / 15.8
NL Dutch 52.6 / 24.1 โ€” 44.3 / 17.5
TH Thai 57.4 / 58.6 โ€” 43.5 / 41.9
UK Ukrainian 51.0 / 24.0 โ€” 41.7 / 18.0
EL Greek 38.1 / 13.3 โ€” 41.1 / 14.6
HI Hindi 48.4 / 31.5 20.5 / 6.0 38.2 / 20.4

License & Attribution

Licensed under Apache-2.0 (LICENSE.txt, ATTRIBUTIONS.md), derived from qvac/TranslatePsy-AfriSLM-2B by Tether AI Research (Milan Gritta, Patrik Lambert, Jihye Back, Amril Nazir) and Qwen/Qwen3.5-2B.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for melodyleonard/AfriSLM-2B-2bit-kronos

Finetuned
Qwen/Qwen3.5-2B
Finetuned
(713)
this model