A 1.5B LoRA fine-tune of unsloth/Qwen2.5-1.5B-Instruct that turns it into an agentic coding model for the
Synoema programming language — it writes Synoema, type-checks it,
runs it, searches a corpus, and self-corrects on errors, all through MCP tools.
🏆 Result: 100% (28/28) on the Synoema agentic tool-use benchmark
Scored on the corrected agentic harness: the model is driven turn-by-turn (generation
stops at <|im_end|>), and real tool results are injected between turns — actual
sno check / sno run output from the live Synoema compiler, never mocked. A task only passes
if the model genuinely completes it end-to-end (e.g. multi-write self-correction: write broken
code → observe the type error → rewrite a valid fix → type-check passes).
Capability
Tasks
Pass
Write + typecheck + run
TU1–TU3, TU5, TU10
✅
Search → write → run
TU6, TU9, TU20
✅
Multi-write self-correction (if/else → ternary)
TU4, TU13
✅
Language features (ADT, HOF, pattern match, cons)
TU11, TU14–TU19, TU23, TU29
✅
List comprehensions
TU12, TU26
✅
Nested ternary (fizzbuzz)
TU22, TU30
✅
Total
28
28/28
What is Synoema?
Synoema is an LLM-native programming language and runtime designed so
that models can write it reliably:
BPE-aligned operators — every operator maps to exactly one cl100k_base token.
Ternary instead of if/else — ? cond -> a : b (nestable).
GBNF grammar for constrained decoding (structural-correctness guarantee).
Cranelift JIT + WebAssembly compile targets.
MCP server exposing file_write, file_read, sno_typecheck, sno_run, search_corpus.
Contract annotations (requires / ensures) for formal verification.
Model details
Property
Value
Base model
unsloth/Qwen2.5-1.5B-Instruct
Parameters
1.5B
Method
QLoRA (4-bit NF4 + LoRA), merged to fp16 for GGUF
LoRA
r=16, alpha=32
Sequence length
1024
Epochs / cycle
3
Training corpus
~18k tool-use + codegen examples — every example passes sno check + sno run
Cycle
C12 (sequential "carousel": each cycle warm-starts from the best previous adapter, then trains on the corpus plus targeted examples for the prior cycle's failures)
Hardware
AMD RX 7900 GRE 16GB (ROCm + unsloth)
GGUF files (llama.cpp / Ollama / LM Studio)
File
Quant
Size
Notes
synoema-coder-1.5b-tools-v12.Q4_K_M.gguf
Q4_K_M
940 MB
smallest, recommended for local use
synoema-coder-1.5b-tools-v12.Q8_0.gguf
Q8_0
2 GB
near-lossless
synoema-coder-1.5b-tools-v12.f16.gguf
F16
3 GB
full precision
# llama.cpp
llama-cli -hf delimitter/synoema-coder-1.5b-tools-v12 --hf-file synoema-coder-1.5b-tools-v12.Q4_K_M.gguf -p "Write quicksort in Synoema to src/qs.sno and run it."# Ollama
ollama run hf.co/delimitter/synoema-coder-1.5b-tools-v12:Q4_K_M
Prompt format is ChatML. The system prompt used at training/eval:
<|im_start|>system
You are sno-code, a Synoema coding agent. Use tools to write and verify code.<|im_end|>
<|im_start|>user
Write `square x = x * x` with `main = square 9` to src/square.sno, typecheck and run it.<|im_end|>
<|im_start|>assistant
The model emits OpenAI-style tool_calls for file_write, sno_typecheck, sno_run,
file_read, search_corpus; feed real tool results back as tool turns.
Synoema language quick reference
maxOf x y = ? x > y -> x : y -- ternary (NO if/then/else)
fact 0 = 1 -- pattern matching
fact n = n * fact (n - 1)
evens xs = [x | x <- xs, x % 2 == 0] -- list comprehension
sumList xs = foldl (\acc x -> acc + x) 0 xs -- higher-order functions
Direction = North | South | East | West -- ADT
opposite North = South
main = qsort [3 1 4 1 5] -- lists are SPACE-separated
