GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
FableMogger9000
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
FableMogger9000
Model Tree
unsloth/Qwen2.5-Coder-7B-bnb-4bit
Input Modalities
Output Modalities
Supported Functionality
| neume → west | west → neume | |
|---|---|---|
| exact_match | 96.0% | 13.7% |
| pitch_accuracy | 99.1% | 76.5% |
| interval_accuracy | 99.2% | n/a (output is neumes) |
| melodic_equivalence | 97.8% (1.955 / 2.0) | 49.0% (0.981 / 2.0) |
| strict_pass_rate | 96.2% | 16.9% |
Held-out: 4,794 sequences, zero leakage, deterministic scoring (no LLM judge). All metrics shown as percentages; melodic-equivalence is a 0–2 rubric reported as % of its 2.0 maximum.
neume → west transcription (the headline capability). Given a neume sequence and its ison (starting-pitch) anchor, the model produces the exact Western pitch sequence 96% of the time, correctly handling:
oligon_kentema +3, oligon_hypsili +4, and the ypsili combinations up to +7);apli / dipli / tetrapli produce held notes, rendered inline as
<pitch>:<beats> (e.g. C4:5);<|im_end|>
stop-token was used at inference).oligon and petaste are both "+1 step"), so a
rising step is genuinely two-valued and cannot be uniquely recovered from pitches alone. Judge
this direction by positional pitch accuracy (76%), not exact match.The same class of model scored ~10% positional accuracy on the real corpus and 96% exact on correct-by-construction data. That contrast isolates the earlier failure as a data-labeling problem, not a model-capability problem: when the neume↔pitch pairs actually correspond, a small fine-tuned model learns the grammar almost perfectly.
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelbase = "unsloth/Qwen2.5-Coder-7B-bnb-4bit"tok = AutoTokenizer.from_pretrained(base)model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")model = PeftModel.from_pretrained(model, "YOUR_USERNAME/byzantine-synthetic-grammar-lora")prompt = ("Transcribe this Byzantine neume sequence (6 neumes) to Western staff pitches:\n""Mode 1\nIson: D4\noligon_kentema petaste elaphron petaste ison oligon_hypsili")msgs = [{"role": "system", "content": "You are a Byzantine chant notation assistant."},{"role": "user", "content": prompt}]ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)out = model.generate(ids, max_new_tokens=96,eos_token_id=tok.convert_tokens_to_ids("<|im_end|>"))print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))# -> Mode 1 / Ison: D4 / G4 A4 F4 G4 G4 D5 (example shape)
Input format. Line 1: an instruction naming the token count. Line 2: Mode X. Line 3:
Ison: <pitch> (the starting-pitch anchor — pitches are relative, so this is required). Line 4:
the space-separated neume (or, for west→neume, pitch) tokens.
ison 0 · oligon/petaste +1 · apostrophos −1 · elaphron −2 · elaphron_apostrophos −3 ·
chamile −4 · oligon_kentema +3 · oligon_hypsili +4 · ypsili_left_oligon +5 ·
ypsili_kentima_oligon +6 · ypsili_over_kentima_oligon +7 (octave). Duration signs:
apli (2 beats) · dipli (3) · tetrapli (5). Breath/barline signs are no-ops on pitch
(neume→west only).
unsloth/Qwen2.5-Coder-7B-bnb-4bit (4-bit), LoRA via PEFT, response-only loss.Educational / research demonstration that supervised fine-tuning teaches a small model a notation grammar that prompting a frontier model did not reliably perform. Not a production transcription tool for real manuscripts, and not a source of musicological ground truth for microtonal or chromatic repertoire.
Apache-2.0 (adapter). The base model retains its own license.