JonusNattapong/lfm3-coding-tool-use-v4.2-pilot

LFM3 coding/tool-use v3 pilot — draft model card

Private staging package. Not a production coding-agent model.

This experiment trains a new LoRA on 32 individually authored coding/tool-use conversations. It starts from the pinned abliterated derivative rather than continuing the earlier adapter with unresolved dataset provenance.

The base revision, actual training configuration, dataset hashes and measured results are retained in the experiment artifacts. Results must be read from the verified report, not inferred from the training loss.

Intended use

Study next-action tool selection and final-answer behavior in a small text-only model. The model may emit Liquid-native tool calls; the included converter validates and maps these to OpenAI-shaped objects. It does not provide a complete serving endpoint or a sandboxed coding agent.

Limitations

• Pilot-size training and evaluation sets; no reliable generalization estimate.
• Four constrained closed-loop repair tasks were tested: base passed 4/4 and v3 passed 3/4. These are too few for general conclusions; no larger-model baseline was evaluated.
• No verified OpenClaw, Hermes-Agent or Grok bot integration.
• No claim of preserved audio/vision abilities after text-only LoRA.
• Tool outputs are untrusted; a host must enforce filesystem containment, execution limits, secret protection and approval of consequential actions.
• Raw reasoning output is retained privately for reproducibility and must be reviewed separately before any publication.

License and provenance

The base is subject to its published LFM license terms and derivative provenance. Do not label this adapter unrestricted commercial-use. The authored dataset's distribution license is not yet assigned by its owner. Public model and dataset release remains blocked on final license/notice/provenance review.

Do not treat a Hub license label alone as commercial clearance.

Measured pilot results

On 16 coding next-action probes, v3 had 15/16 correct call/no-call decisions, 10/10 exact expected tool calls and 1/6 unnecessary calls. The pinned base had 15/16 decisions, 9/10 exact expected calls and 1/6 unnecessary calls. Actual four-task sandbox repair favored the base in this run (4/4 versus 3/4). No statistical significance or broad capability gain is claimed.

Adapter and reproduction

Base revision: 22c21265617da773c3cb55fb904f517b395a57ab. Fresh LoRA: rank 16, alpha 32, dropout 0.05, LR 5e-5, 3 epochs, seed 42, microbatch 1 and gradient accumulation 2. The run completed 48 optimizer steps. Adapter weight SHA-256: c6665fa6188883de9489b0f3121caf5bd55e184044cc677739f77aab0c6daa02.

This is a PEFT adapter, not a standalone merged base model. Loading the base requires trusted model-specific code. Review that code and keep the pinned revision when reproducing the experiment. Exact software versions and dataset hashes are included in training_config.json.