Flagship · current
Pro 2508
Chains eleven grounded tools, reasons in full, and sings in verse. The deepest Quenya–Sindarin transformations live here.
The workbench
A language kept alive,
word by word.
Eldalambe is a family of small, fine-tuned models that speak Neo-Quenya and Neo-Sindarin — words the lexicon attests it quotes with sources, words it lacks it coins with roots laid bare, and what cannot be built honestly it refuses. Every reply is checked against the Eldamo snapshot before it reaches you.
It is also a proof of method: a blueprint any small-language community can copy — corpus, grounding, small fine-tunes — to keep its own tongue alive on a modest budget.
The four weights
One lexicon, one grammar engine, one verification loop — four capacities. Select a model to open it in the chat.
Lite · current
Lite 1.2B
Short translation and lexicon lookup, one-shot Elvish — light on its feet without giving up the canon.
Vision · current
OCR 3B
Reads handwriting from a Tengwar image — returns the script itself and a romanized reading.
Previous generation
Pro 1304
The earlier adapter of the same base — kept as a fallback and a regression anchor for evaluating 2508.
Why it exists
Elvish is the demo.
Your language is the point.
A living language needs speakers, a corpus, and tools that keep it honest. Big models hallucinate rare words; big budgets lock small communities out. Eldalambe is our answer to both — a method for keeping a small community language alive with models small enough to be cheap, open and local.
-
I
Start from a canon
Everything rests on a versioned lexicon snapshot — here, Eldamo: attested forms, glosses, roots, sources. For your community it is your dictionary, your texts, your recordings. No canon, no claim.
-
II
Ground every answer
The models do not answer from memory alone. Lookup, coin, compose, pronounce, read — each is a tool call into the canon. No tool call, no claim. When the lexicon holds no word, the model coins one: attested roots, phonotactics respected, the result labelled a coinage — the way living languages grow — never passed off as attested. What cannot be built honestly, it refuses.
-
III
Fine-tune small, evaluate hard
1.2B–3B parameter adapters, trained on canon-derived data — cheap to train, cheap to serve, happy on a single GPU. The previous generation is kept alive as a regression anchor, so every new adapter must beat it to ship.
-
IV
Hand it to the community
Chat, tools, speech and script — behind a simple invite, with credits instead of keys. Coinages are proposals the community accepts or declines, so the lexicon grows the way a living language does. A community-run instance of its own tongue: that is the whole idea. Elvish proves it works; the next language can be yours.
“A language that is not used, is not lived.” The tools on this site — the chat, the lexicon, the voice, the script — are the same ones a community language needs on day one. The Elvish is the demonstration; the method is the product.
The evidence
What is actually running behind this page — adapters, datasets, and the verification chain.
~221M
tokens in the Pro SFT corpus
282K + 217K
training examples · Pro + Lite
r=128
LoRA rank on every adapter · α=181
92.7%
honest refusals on elvbench — our benchmark, 956 probes
Adapters
Small bases, real lineage
- Pro 2508 — LiquidAI LFM2.5-2.6B + LoRA (r=128, α=181) · lineage
2.6B → r2 → v3-3ep - Lite 1.2B — LFM2.5-1.2B-Instruct + LoRA · 4-tool single-turn
- OCR 3B — LFM2.5-VL-3B + LoRA · handwritten Tengwar (PUA → latin)
- Pro 1304 — prior adapter, kept as regression anchor
Training
Two stages, one canon
- QLoRA SFT on canon-derived sets —
eldalambe-sft,eldalambe-lite-sft,eldalambe-vl-sft - GRPO / RLVR — exact-match rewards on attested morphology & Tengwar transcription (
verified/rlvr_prompts.jsonl) - Grammar guard + provenance annotation runs alongside every answer
Evaluation
Scored, not vibes
- elvbench & elvbench-chain — tool-chained morphology & translation
- horo-verify — hour-by-hour regression harness
- Every adapter must beat the previous generation to ship — 1304 exists to be beaten
- General skills are measured too — and traded, openly: MMLU and GSM8K give way to canon fidelity, because the product is a specialist that invents only on purpose — labelled coinage, never silent fabrication
Grounding
The loop you can watch
- 11 grounded tools on Pro: lookup, coin, compose, pronounce, OCR, clock…
- Answers annotated attested / derived / neologism, unverified flagged
- Speech via G2P-tuned Kokoro; script via Tengwar renderer
Four weights, one canon
Every Eldalambe model is trained, evaluated, and served against the same Eldamo snapshot. The differences are of capacity and role — never of what counts as a real word.
| Model | Size | Role | Status |
|---|---|---|---|
Pro 2508 eldalambe-2.6b | 2.6B | 11-tool chained reasoning, full composition | current |
Pro 1304 eldalambe-1304 | 2.6B | Prior adapter, fallback and regression reference | previous |
Lite 1.2B eldalambe-lite | 1.2B | Compact grounded classifier, 4 tools | current |
OCR 3B eldalambe-ocr | 3B | Tengwar vision model (LFM2.5-VL-3B + LoRA) | current |
11grounded tools on Pro
4grounded tools on Lite
3faces: Quenya, Sindarin, Tengwar
1lexicon — Eldamo, versioned
elvbench — our own measure
The benchmarks this project is actually built around: lexicon grounding, coinage honesty, translation and tool-chaining, scored against the canon itself.
| Probe | What it asks | Pro 2508 |
|---|---|---|
| elvbench · honesty | 956 impossible concepts — does the model refuse rather than invent? | 92.7% |
| elvbench · suite | full run: morphology, translation, coinage, refusal — 1,656 items | scored per release |
| elvbench-chain | multi-tool chains: inflect → verify → render script | scored per release |
| horo-verify | hour-by-hour regression harness between releases | gate to ship |
The rule every adapter lives by: it must beat the previous generation on these — Pro 1304 exists to be beaten.
General skills, for the record
lm-eval-harness, Pro 2508 adapter vs its own base (LFM2.5-2.6B), 2026-08-17. A specialist trades breadth for canon fidelity — the full picture, not just the wins.
| Benchmark | Base | Pro 2508 | Δ |
|---|---|---|---|
| IFEval · instruction strict | 14.1 | 38.3 | +24.2 |
| IFEval · instruction loose | 14.1 | 43.0 | +28.9 |
| HellaSwag · acc_norm | 61.3 | 51.3 | −10.0 |
| MMLU · 5-shot | 65.2 | 38.9 | −26.3 |
| GSM8K · strict | 71.3 | 7.5 | −63.8 |
Instruction-following sharpened (+24.2); arithmetic and trivia given up. A language specialist is the product — and the coinage tool keeps it growing: for what the lexicon lacks, the model proposes new words from attested roots, always labelled, always reviewable.
Speak with Eldalambe
Begin a conversation with the Eldalambe.
Grounded answer by grounded answer — every word traced to the lexicon.
Credits are running low. A top-up keeps the conversation flowing — chat costs 20 ✦ per message.
Grounded tools
Each tool calls a lexicon or grammar function inside the model — no tool call, no claim.
Lookup
Canonical dictionary entry for a Quenya or Sindarin word — the attested form, its root, and the source, all in one glance.
Coin
Coin a word the lexicon does not yet hold — attested roots, phonotactics respected, etymology laid bare. Marked as a coinage rather than canon: this is how the language grows.
Pronounce
Hear the Elvish spoken aloud with the G2P-tuned voice. Only Elvish lines are voiced; the English gloss stays silent.
Compose
A short poem in *italics*, with one English meaning beneath each line — metre and meaning kept in balance.
Verify · provenance
Ask a question and see the full word-by-word provenance: attested, derived, or neologism — with the lexicon sources each word rests on. This is the grounding loop, made visible.
OCR · Tengwar
Reads native Tengwar from an image and returns the script itself, along with a romanized reading.
Your account
Credits, usage, and billing. Every member starts with 1,000 credits.
E
—
—
—
Credits remaining
0
—
Usage log
| Action | Cost | Model | When |
|---|---|---|---|
| No usage yet. | |||
Admin
Manage members, credits, and invites.
Members
| User | Credits | Role | Action | |
|---|---|---|---|---|
| Loading… | ||||
Invites
| Code | Used | Max |
|---|---|---|
| Loading… | ||