Cookbook - Personal Code Specialist

Cookbook - Personal Code Specialist

What you get:

• A per-repo corpus extraction flow.
• A posttrainllm checkpoint served as an OpenAI-compatible local coding model.
• Continue.dev and Aider config snippets.
• A small benchmark template for repo-pattern completions.
• A clear limit: this is not GPT-4 for general coding.

What this is not: an editor plugin. Continue.dev and Aider already know how to call OpenAI-compatible local servers; posttrainllm only has to serve the model.

1. Build A Repo Corpus

The example script walks the current repo, keeps source/doc files, and emits JSONL rows with path and text. Run:

examples/repo-specialist-cli/run.sh corpus .

Output:

.tinygpt/repo-specialist/corpus.jsonl

2. Train Or Fine-Tune

Use the corpus as the domain data for a specialist run. The exact command depends on the base model you are using; keep it small and eval-driven:

posttrainllm sft \
  --base ~/.cache/posttrainllm/runs/huge-base-v1/huge-base-v1.tinygpt \
  --corpus .tinygpt/repo-specialist/corpus.jsonl \
  --steps 2000 \
  --out .tinygpt/repo-specialist/model.tinygpt

If you already have a checkpoint:

export TINYGPT_MODEL=.tinygpt/repo-specialist/model.tinygpt

3. Serve It

posttrainllm serve "$TINYGPT_MODEL" --host 127.0.0.1 --port 8080

4. Continue.dev

Continue’s OpenAI provider accepts apiBase. Add this to ~/.continue/config.yaml:

name: posttrainllm Local Specialist
version: 0.0.1
schema: v1
models:
  - name: posttrainllm Repo Specialist
    provider: openai
    model: posttrainllm
    apiBase: http://127.0.0.1:8080/v1
    apiKey: not-needed
    roles:
      - chat
      - edit
      - apply
    capabilities:
      - tool_use

Generate the snippet:

examples/repo-specialist-cli/run.sh continue

5. Aider

Aider can connect to an OpenAI-compatible endpoint with environment variables:

export OPENAI_API_BASE=http://127.0.0.1:8080/v1
export OPENAI_API_KEY=not-needed
aider --model openai/posttrainllm

Generate the snippet:

examples/repo-specialist-cli/run.sh aider

6. Benchmark

Create a held-out file of repo-specific prompts:

{"prompt":"Add a route matching the existing Astro page style.","expected_contains":"frontmatter"}
{"prompt":"Write a Swift CLI subcommand using the existing run(args:) pattern.","expected_contains":"static func run(args:"}

Then run:

posttrainllm run-bench \
  --model "$TINYGPT_MODEL" \
  --tasks custom-code-patterns \
  --limit 50 \
  --out docs/artifacts/repo-specialist-bench.jsonl

Record:

Honest Limitations

• This helps with local idioms, naming, and repetitive patterns. It will not become a frontier code model.
• Bad repo data makes a bad specialist. Exclude generated files and vendored dependencies.
• Keep the first eval tiny and concrete before scaling training.

See also: smolagents and Pydantic AI.