Level 8 project

Adapt a Small LLM

Start from the canonical project files, validate the result, and keep the evidence you need to explain what you built.

Start here

Launch lesson: Fine-Tuning Integration Workshop

Prerequisite project: Reliable LLM Workflow

Download project files→View on GitHub

Goal

Build a reproducible model-adaptation package that can answer:

Why adapt?
→ which data?
→ which split/template?
→ which base?
→ which LoRA/PEFT configuration?
→ which adapter?
→ how would RLHF-style and DPO-style preference optimization differ here?
→ which measured signals are only proxies?
→ what independent evaluation supports the claim?
→ what improved?
→ what regressed?
→ does the run meet the declared release rule?

Task

lora_parameter_count
group_split
release_decision
validate_run

Validation

Run these commands from the downloaded Project folder or the public materials repository root.

python projects/tests/l08/validate_submission.py \
projects/starters/l08/adaptation.py \
path/to/your/adaptation-run.json

Rubric

Score each criterion from 0–4. A submission should score at least 16/20 overall and must not score 0 on data boundaries, adaptation lineage, or evaluation.

Criterion	4 — Strong evidence	3 — Meets	2 — Partial	1 — Weak	0 — Missing
Adaptation decision	Measured behavioral gap; prompting/retrieval/application alternatives considered; fine-tuning choice is evidence-based	Clear fine-tuning rationale with small omissions	Rationale exists but alternatives/evidence are shallow	Fine-tuning chosen mainly because it is available	No adaptation rationale
Data and split integrity	Provenance, grouping, dedup/conflict policy, formatting version, and train/eval separation are reproducible	Data pipeline is reproducible with minor gaps	Some provenance/split evidence missing	Dataset dump with weak leakage controls	No trustworthy data boundary
PEFT/LoRA implementation and lineage	Rank/count/targets/base compatibility are correct and all artifacts identify their parents	Correct adapter implementation/config with minor gaps	Adapter works but identity/count evidence incomplete	Adapter exists with unclear base/config	No coherent adaptation artifact
Before/after, retention, and preference-context evaluation	Paired target and retention suites, predeclared thresholds, regression IDs, and release decision are supported; RLHF-style vs DPO-style routes are distinguished; preference/reward signals are explicitly separated from independent quality evidence	Complete evaluation and preference-method explanation with minor gaps	Target evaluation present but retention, route distinction, or proxy-vs-independent evidence is weak	Cherry-picked generations, training loss, or preference/reward score treated as sufficient quality evidence	No held-out adaptation evaluation
Documentation and reproducibility	Training record, model card, limitations, debug trace, commands, and optional live-run differences are explicit	Reproducible package with small omissions	Important versions/limitations missing	Mostly screenshots or output dumps	No reproducible package

Objective checks

The reduced validator checks:

LoRA parameter-count arithmetic;
group-safe train/validation splitting;
target-gain and retention-drop release rules;
zero critical-regression requirement;
base/adapter lineage;
required dataset/adaptation/evaluation identities;
a structured RLHF-style vs DPO-style route contract and selected optimization route;
proxy metric identity when a preference route is used;
independent-evaluation IDs linked to the recorded target and retention metrics and separated from the proxy metric;
model-card limitations;
intentional failure rejection.

Human review

Human review should inspect whether:

fine-tuning was the right intervention for the declared gap;
target data actually represents desired behavior;
RLHF-style reward-model/policy optimization and DPO-style direct preference optimization are not presented as the same route;
preference, reward, or SFT evidence is not overclaimed as truth or broad quality;
target gains are worth any measured regressions;
limitations state where evidence does not apply.

Assessment guardrail

A live GPU fine-tune is optional for repository acceptance. If provided, it extends the evidence but does not replace the deterministic reduced path or the requirement for before/after retention evaluation.

← Back to all projects