enseven/lfm-2.5-think-code-GGUF

LFM2.5-1.2B-Thinking KodCode fine-tune — GGUF collection (negative result)

The exact GGUF artifacts used as the four model rows in a sealed 128-task HumanEval+ evaluation (greedy decoding, EvalPlus v0.1.10, pinned llama.cpp runtime). Published so the evaluation is reproducible from the Hub alone. The fine-tune regressed coding ability versus the base model — see the adapter model card for the full result and the recommendation to use the base model.

Artifacts

File	Size (bytes)	SHA-256	Role in the evaluation
lfm2.5-1.2b-thinking-original-bf16.gguf	2,343,326,656	2e948214bc4f2af1990039946cee14ad1525ed41108d3fb77a4ec06c7ceec781	Baseline row (upstream base model, no fine-tune)
lfm2.5-1.2b-thinking-kodcode-bf16.gguf	2,343,326,048	3e39c292be1f2b2740c3e05a2cb9081260439033e1b05f601117c72c8a73b8e6	Fine-tune merged BF16 reference row
lfm2.5-1.2b-thinking-kodcode-q8_0.gguf	1,246,253,408	3e2223fa4baf49896f1a58d0b73e02f5143b745d8deb71a01a6038b7994ce394	Pilot + sealed quantization row
lfm2.5-1.2b-thinking-kodcode-q6_k.gguf	962,842,976	d1a14fa33fba75118c8262827bfcadc89d0c549d628ac243014df81a14d8f331	Pilot + sealed row; passed the predeclared deployment gate (−1.59 pp vs merged BF16, no new failure class)

Q5_K_M / Q4_K_M / Q2_K exist locally but were never benchmark rows and are not published.

Provenance

• Converted from the Hugging Face checkpoints with the pinned llama.cpp toolchain (commit 3018a11e79e489b657dbb77c95694889ccff92df), offline, in restricted read-only containers.
• Each quantization was produced independently from the validated BF16 reference — never requantized from a quantized file.
• Conversion parity was gated per artifact (tokenizer output and first greedy token parity against the source checkpoint).
• Evaluated at 4,096-token context, F16 K/V cache, all layers on GPU, --fit off, parallel 1.

Usage

llama-server --model lfm2.5-1.2b-thinking-kodcode-q6_k.gguf \
  --ctx-size 4096 --n-gpu-layers all --flash-attn on \
  --cache-type-k f16 --cache-type-v f16 --fit off --parallel 1 --offline

Prompting note from the evaluation: this fine-tune was trained to answer code requests directly without a reasoning trace. That formatting is clean, but on this benchmark it reduced correctness versus letting the base model think — see the model card above before choosing an artifact.

License

CC BY-NC 4.0 — inherited from the KodCode-V1-SFT-R1 training data used to produce the fine-tune rows. The upstream base model remains subject to its own Liquid AI license terms.