LFM-2.5-8B-1B Hermes FT
This repo contains Hermes/tool-use fine-tuned variants of LiquidAI/LFM2.5-8B-A1B.
The current release candidate is iter13_llamacpp_chat_fixed. Earlier GGUF artifacts were withdrawn because they overfit tool routing and regressed normal chat by stopping too early. The iter13 repair adds a narrow chat-retention LoRA pass on top of the fixed Hermes tool-router model, then regenerates MLX and GGUF quants from the repaired fused checkpoint.
Runtime requirements
- MLX releases preserve
tool_parser_type: "pythonic".
- llama.cpp GGUF releases are intended to be served with the LFM chat template,
--jinja, and a 64K context smoke target.
- Tested live server alias:
sjakek/LFM-2.5-8B-1B-hermes-ft:Q6KXL.
Available artifacts
gguf/LFM-2.5-8B-1B-Hermes-Tuned-Q8KXL.ggufgguf/LFM-2.5-8B-1B-Hermes-Tuned-Q6KXL.ggufgguf/LFM-2.5-8B-1B-Hermes-Tuned-Q5KXL.ggufgguf/LFM-2.5-8B-1B-Hermes-Tuned-Q4KXL.ggufmlx/8bit/mlx/6bit/adapters/iter13_llamacpp_chat_retention_r8/evals/iter13_*
Validation summary
All listed iter13 variants passed:
- Normal-chat regression:
10/10
- Fixed-Hermes tool-router suite:
43/43
- Structured tool-call cases:
28/28
- No-tool false positives:
0/10
- Text tool-call leaks:
0
The normal-chat eval includes plain chat, tools-available no-tool chat, short factual answers, and multi-sentence explanations. The tool suite uses the fixed Hermes browser, terminal, file/search/write, no-tool, and tool-result finalization cases.
Training summary
- Semi-full-gradient grouped MoE expert training over the Hermes trace corpus.
- Fixed-Hermes contrastive router LoRA repairs for structured pythonic tool calls.
iter12 chat-retention repair to fix MLX normal-chat early stopping.iter13 llama.cpp-targeted chat-retention repair after BF16 GGUF exposed shorter completions than MLX.
The GGUFs were regenerated from a dequantized fused safetensors source, then quantized from the BF16 GGUF parent. They are named Q*KXL as Hermes-tuned mixed-precision KXL targets.
