Fine-Tuning Run Report: [Run name or ID]

Report for one fine-tuning run of a pretrained model: objective, base model and licence, data, method and hyperparameters, compute, training curves, evaluation against the baseline, safety evaluation, failure analysis, decision, artefacts and reproducibility. Based on common machine learning practice. Replace every [bracketed] value and every row marked Example. Worked examples use a LoRA fine-tune of an open-weight 8B instruction model for support ticket summarisation.

1. Objective and success criteria

Objective. Produce ticket summaries of at most [80] words that keep the customer's issue, steps already taken and requested outcome, in the house style, at lower cost than the current [hosted large model].

Why fine-tuning. Prompting the base model reached [0.71] on the rubric score with a [1,900]-token prompt; a smaller tuned model is expected to match quality at about [one fifth] of the cost.

2. Base model and licence

3. Data

3.1 Sources

3.2 Filtering and cleaning

• Removed tickets with fewer than [3] messages or summaries under [10] words: [-3,900].
• Removed near-duplicates (MinHash similarity > [0.9]): [-2,100].
• Removed items flagged by PII scanner after pseudonymisation: [-410].
• Decontamination: removed any ticket whose ID or text appears in the evaluation sets: [-35].

3.3 Size and splits

4. Method and hyperparameters

Method: [LoRA supervised fine-tuning on chat-formatted pairs, loss on assistant tokens only].

With LoRA the weight update is a low-rank product $\Delta W = BA$ with $B \in \mathbb{R}^{d \times r}$ and $A \in \mathbb{R}^{r \times k}$, scaled by $\alpha / r$, so only $r(d + k)$ parameters per matrix train.

5. Compute

6. Training curves

[Description and link to plots.] Example: "Training loss fell from 1.92 to 0.88 in the first epoch and to 0.71 by the end. Validation loss reached its minimum 0.83 at step 1,650 (epoch 2.5) and rose slightly to 0.85 at the end, so the checkpoint at step 1,650 was selected. No loss spikes occurred; gradient norm stayed below 1.0 after warm-up."

7. Evaluation against baseline

8. Safety evaluation

9. Failure analysis

10. Decision

Decision: [Promote to shadow / Promote / Iterate / Reject]. Example: "Promote to a 2-week shadow deployment. SC-01, SC-02 and SC-04 are met; SC-03 shows no regression. Next run will address the missing-outcome failure group."

11. Artefacts and reproducibility

Rerun command: [train --config configs/ft-2026-031.yaml --seed 42].

Standard and licence note

Based on common industry practice for documenting fine-tuning experiments of machine learning models. Original text written for this collection; example numbers are illustrative. Check the licences of the base model and the training data for your own use. Template licence: MIT.