# LFM2.5 Hermes Tool-Trace LoRA Run Report

## Outcome

Completed a local MLX LoRA fine-tune for `LiquidAI/LFM2.5-8B-A1B-MLX-bf16` on Hermes-style tool traces. The final adapter reloads, the fused MLX checkpoint loads, and the fixed mini-suite shows a small tool-format improvement over the baseline GGUF server.

GGUF export is blocked in `mlx-lm` for this architecture:

```text
ValueError: Model type lfm2_moe not supported for GGUF conversion.
```

## Environment

- `mlx`: 0.31.2
- `mlx-lm`: 0.31.3
- `datasets`: 4.8.5
- `transformers`: 5.9.0
- `huggingface_hub`: 1.17.0
- Python: `/opt/homebrew/bin/python3` via `lfm25_hermes_ft/venv`

## Dataset

Source dataset: `DJLougen/hermes-agent-traces-filtered`, `default/train`.

Initial 16K artifact:

- Input rows: 3,679
- Accepted rows at 16K cap: 1,583
- Split: 1,431 train, 90 valid, 62 test
- Rejected as over 16K: 2,077
- Rejected as malformed/unbalanced: 19

The requested 16K training pass was attempted and failed with Metal OOM. A deterministic 12K view was created from the accepted 16K artifact:

- Train: 880 kept, 551 dropped from 16K artifact
- Valid: 47 kept, 43 dropped from 16K artifact
- Test: 36 kept, 26 dropped from 16K artifact

## Training

Final training config:

- Model: `LiquidAI/LFM2.5-8B-A1B-MLX-bf16`
- Data: `lfm25_hermes_ft/datasets/hermes_filtered_text_12k`
- LoRA rank: 16
- LoRA layers: 16
- Batch size: 1
- Gradient accumulation: 1
- Max sequence length: 12,000
- Iterations: 880
- Learning rate: `1e-5`
- Validation batches: 1
- Gradient checkpointing: enabled
- Seed: 42

Final training line:

```text
Iter 880: Val loss 0.465, Train loss 0.278, Tokens/sec 974.875, Trained Tokens 7501092, Peak mem 50.948 GB
```

## Evaluation

Fixed six-case Hermes mini-suite:

- Single weather tool call
- Web search tool call
- Calculator tool call
- Tool-response finalization
- No-tool natural response
- Malformed-tool robustness

Results:

| Lane | Passed | Parseable tool calls |
| --- | ---: | ---: |
| Baseline GGUF server | 3 / 6 | 0 |
| MLX LoRA adapter | 4 / 6 | 1 |

The adapter produced a valid Hermes XML/JSON tool call for the web-search case:

```xml
<tool_call>
{"name": "web_search", "arguments": {"query": "Liquid AI LFM2.5 release notes"}}
</tool_call>
```

It did not yet reliably force tool calls for weather or calculator prompts.

## Artifacts

- Final adapter: `lfm25_hermes_ft/checkpoints/hermes_lora_r16_12k/adapters.safetensors`
- Adapter checkpoints: `lfm25_hermes_ft/checkpoints/hermes_lora_r16_12k/0000100_adapters.safetensors` through `0000800_adapters.safetensors`
- Fused MLX checkpoint: `lfm25_hermes_ft/artifacts/hermes_lora_r16_12k_fused_mlx`
- Eval JSON: `lfm25_hermes_ft/evals/hermes_tool_eval.json`
- Training log: `lfm25_hermes_ft/logs/train_lora.log`
- GGUF blocker log: `lfm25_hermes_ft/logs/fuse_export_gguf.log`
- OOM evidence:
  - `lfm25_hermes_ft/logs/train_lora_oom_val4_with_server.log`
  - `lfm25_hermes_ft/logs/train_lora_oom_val1_no_server.log`
  - `lfm25_hermes_ft/logs/train_lora_oom_gradaccum1_no_server.log`

## Status

This run is useful as a retained MLX adapter/fused checkpoint, but it is not llama.cpp-ready because GGUF export for `lfm2_moe` is unsupported by the installed `mlx-lm` exporter.