lfm2.5-230m-fable-5

LFM2.5-230M Fable-5 GGUF

Fine-tuned GGUF release of LiquidAI/LFM2.5-230M on Glint-Research/Fable-5-traces.

Files

• lfm2.5-230m-fable-5-f16.gguf — highest quality, largest file
• lfm2.5-230m-fable-5-q8_0.gguf — high quality, smaller
• lfm2.5-230m-fable-5-q4_k_m.gguf — best default for local inference

Training

• Base model: LiquidAI/LFM2.5-230M
• Dataset: Glint-Research/Fable-5-traces
• File used: fable5_cot_merged.jsonl
• Method: PEFT LoRA SFT
• Max sequence length: 4096
• Epochs: 1
• LoRA rank: 32
• LoRA alpha: 64
• LoRA dropout: 0.05
• Precision: FP16 base model, FP32 LoRA trainable weights
• Hardware: Google Colab T4
• Format: Chat template system/user/assistant, preserving Fable context -> completion

Final training loss samples

• step 555: 1.7037
• step 560: 1.5968
• step 565: 1.6435
• step 570: 1.6109
• step 575: 1.6589
• step 580: 1.6439

Evaluation

We evaluated AKMESSI/lfm2.5-230m-fable-5:F16 against the original base model, LiquidAI/LFM2.5-230M-GGUF:BF16, using local llama.cpp server inference.

These are not official leaderboard submissions. They are lightweight local evaluations intended to compare the fine-tuned model against the base model under the same prompts, decoding settings, and hardware setup.

Summary

The Fable-5 fine-tune improves repository-context code continuation on RepoBench-C-lite Python, while mostly preserving the base model's generic function-calling behavior on BFCL-lite Simple.

---

RepoBench-C-lite Python

RepoBench-C-style next-line code completion was used to evaluate repository-context code continuation. We sampled 100 examples each from python_if, python_cff, and python_cfr, for 300 total examples.

Compared with the base model, the Fable-5 fine-tune improved:

• Exact match by +4.33 percentage points
• Prefix match by +4.67 percentage points
• Edit similarity by +3.32 points

Breakdown by config:

---

BFCL-lite Simple

We also ran a local BFCL-lite Simple function-calling evaluation over 400 examples as a generic tool-calling control.

The fine-tuned model preserves most of the base model's generic function-calling behavior, but does not improve BFCL-style API-schema-to-JSON calling. This is expected because the training data consists of coding-agent traces rather than clean function-calling examples.

---

CodeXGLUE Line Completion Python

We ran a 1,000-example local CodeXGLUE line-completion evaluation as a general code-completion control.

This result is effectively neutral. The Fable-5 fine-tune does not materially change general line-completion performance on this setup.

---

CRUXEval-lite

We also tried a 200-example CRUXEval-lite run for Python execution reasoning.

This benchmark was not a good fit for the fine-tuned model. The Fable-5 model often entered explanation or trace-style response mode instead of returning only the exact literal Python value expected by CRUXEval.

---

Interpretation

The Fable-5 fine-tune appears to shift the base model toward coding-agent and repository-context continuation behavior.

It improves RepoBench-C-lite Python next-line completion, while mostly preserving generic function-calling ability on BFCL-lite Simple. The main regression is in exact BFCL-style argument filling, which is not the main target of the Fable-5 trace dataset.

The model is best understood as a tiny coding-agent trace model, not a general-purpose reasoning model or a benchmark-specialized function-calling model.

---

Evaluation Caveats

• These are local lightweight evaluations, not official leaderboard submissions.
• Results were produced with llama.cpp server inference.
• Scores may vary with prompting, decoding settings, quantization level, and benchmark harness details.
• BFCL-lite and RepoBench-C-lite use simplified local scoring scripts rather than official leaderboard infrastructure.
• Only the F16 model was benchmarked here; quantized GGUF variants may differ slightly.

Usage

Recommended local file:

lfm2.5-230m-fable-5-q4_k_m.gguf

Caveats

This model is trained on coding-agent trace telemetry. It may emit tool-call-like actions, shell commands, file paths, or long reasoning-style continuations. Review outputs before executing commands.

The dataset contains coding-agent traces and should not be treated as a clean benchmark or a safety-filtered assistant dataset.

License notes

• Base model: LiquidAI LFM Open License v1.0
• Dataset: AGPL-3.0
• This repo preserves upstream license notices. Check compatibility before commercial or closed-source use.