Fine-tune a coding model on your own agent transcripts (Tinker)
You are helping someone turn the sessions they already ran with a coding agent into training data, and fine-tune an open model on it with Tinker. This skill distils a pipeline that shipped several public models. Follow the steps in order; each script is small, offline unless stated, and prints what it did.
All commands below run from this skill's directory (tinker-ai-skill). Read references/ when a step
needs judgement; the scripts' docstrings (python scripts/<name>.py --help) explain
every flag.
Before you start
- Keep the transcripts. Claude Code deletes transcripts after 30 days by default.
Tell the user to add
"cleanupPeriodDays": 3650to~/.claude/settings.jsonnow. - Keys. Copy
.env.exampleto.envand fill inTINKER_API_KEY(console at tinker.thinkingmachines.ai) andWANDB_API_KEY(wandb.ai/authorize). Never print, commit, or paste these values anywhere; the scripts read.env. - Environment.
uv venv && uv pip install -r requirements.txt(orpip install -r requirements.txt). Python 3.11+. Use that interpreter for every script. - Budget. Training costs tokens x price. Always run a script's default (offline)
mode first; only
--train,sample_checkpoint.pyandexport_checkpoint.pytouch the service. Confirm the estimated cost with the user before launching. - For local serving (steps 8-9). A built llama.cpp checkout recent enough for the model's architecture, disk for the base model plus the merged copy (about 2 bytes per parameter each), and RAM for the largest shard.
The workflow
Step 1. Collect transcripts into a dataset location
python scripts/collect_sessions.py --data-dir ./data # ~/.claude/projects by default python scripts/collect_sessions.py --data-dir ./data --project my-repo --since 2026-01-01
Copies main-session files verbatim (no subagent sidechains), skips tiny sessions,
records sha256s in data/originals/manifest.json. Use --exclude-slug for projects
that must never leave the machine (client work, security research).
Step 2. Clean: normalize, then remove secrets and identity
python scripts/normalize_sessions.py --data-dir ./data python scripts/anonymize_sessions.py --data-dir ./data --identity-file identity.json
normalize_sessions.py turns the raw records into OpenAI-style messages: user text
the person typed, assistant text and plaintext thinking, tool calls with their JSON
arguments, tool results clipped to head+tail. It drops harness-injected text,
sidechains, API-error placeholders, thinking signatures and image bytes, and
verifies every tool call is answered. See references/claude-code-transcripts.md.
anonymize_sessions.py first replaces credentials by value across the whole corpus
(one marker per value), then de-identifies each session with its OWN random
identity: home username, top-level folders under the home directory, project slugs,
and every name/email/handle/org/domain listed in the identity file. Write that file
with the user (their username, full name, emails, GitHub handle, company). The run
fails if any listed string or credential candidate survives. Read the printed list
of detected usernames; pass --keep-username for service accounts that are not a
person. Why per-session randomization and not one fixed fake world:
references/anonymization.md.
Inspect a few anonymized rows with the user before continuing. Only the user can judge whether something is still sensitive.
Step 3-5. Slice to the context length, add compaction context, split train/val
python scripts/build_dataset.py --data-dir ./data --model Qwen/Qwen3.6-35B-A3B --max-len 65536 \ --system-prompt-file assets/system-prompt.md
Pick --model from Tinker's model list and --max-len from that model's context
column (64K base variants; :peft:262144 variants take 256K). The builder measures
every row with the real renderer and tokenizer of that model, so rebuild when you
change models: a tokenizer change moved our rows by 14% and overflowed the budget.
- Windows never split a tool call from its result and never open on a tool message.
- Slice 0 carries the system prompt plus the working directory. Every later slice
carries the same prompt followed by a deterministic
<working_context>block (original request, recent requests, files touched, commands, task state, a trace of the last few messages) and, when the window opens mid-turn, a short "continuation" user message. This is the compaction the model will see in a real long session. Details:references/chunking-and-compaction.md. - Tool definitions are declared on every row. Pass
--tools-filewith the real definitions of the harness the model will serve in; without it, minimal schemas are inferred from usage. Rows without tool definitions regressed tool use for us. - Hybrid-thinking renderers are rebuilt to keep thinking in history (otherwise
earlier turns train on empty think blocks). For Inkling models pass
--effort(default 0.9) and use the same value at serving time. - The split is by whole session, seeded, so rebuilds hold out the same sessions.
Report: data/datasets/<name>.report.json (rows, tokens, loss tokens, sha256s,
rendering provenance). The trainer refuses a dataset whose provenance differs.
Step 6. Fine-tune on Tinker
python scripts/train_tinker.py --data-dir ./data --name qwen3.6-35b-a3b-64k # offline smoke test python scripts/train_tinker.py --data-dir ./data --name qwen3.6-35b-a3b-64k --train # launch
The smoke test renders every row, prints the decoded head of a row (read it: system
prompt, tool declaration, user turn), steps per epoch and total tokens. Multiply
tokens by the model's train price from Tinker's pricing page and tell the user the
estimate before --train.
Defaults: LoRA rank 32, batch 8 rows, 2 epochs, linear decay, eval and checkpoint
every 20 steps, W&B project agent-transcript-sft. Learning rate comes from
Tinker's calibrated get_lr for Qwen/Llama; for Inkling, Nemotron, DeepSeek, Kimi
and GLM you must pass --learning-rate (ranges and evidence in
references/tinker-training.md). When in doubt run two short runs, 1e-4 and 3e-4,
and compare behaviour, not loss.
While it runs: confirm the first metrics reach W&B, watch test/nll at steps 20 and
40 (a dip then a rise means the LR is too hot for this batch; stop and lower it),
and never wrap Tinker calls in your own timeouts or retries.
Step 7. Check behaviour and pick a checkpoint
python scripts/sample_checkpoint.py --data-dir ./data --name qwen3.6-35b-a3b-64k \ --checkpoint tinker://<run-id>/sampler_weights/000100 --prompt "Find and fix the failing test."
Compare base vs checkpoints on prompts from the user's real work. Judge by behaviour (valid tool calls, non-empty thinking, follows the system prompt), not by NLL: a falling loss hid a behaviour regression for us twice. Pick the checkpoint by an eval.
Step 8. Download the LoRA and merge it into the base model
python scripts/export_checkpoint.py --run-dir ./runs/<run> --checkpoint 000100 \ --base-model Qwen/Qwen3.6-35B-A3B --out ./export/my-coder --trust-remote-code
Reads checkpoints.jsonl, downloads the training checkpoint (sampler
checkpoints cannot be downloaded; periodic ones expire after 7 days, so export
before then), merges it with the cookbook's build_hf_model into a standalone bf16
HuggingFace directory, copies generation_config.json, and prints the chat-template
kwarg that restores the training conditioning (earlier-turn thinking kept).
--peft additionally writes a PEFT adapter for vLLM or SGLang. --dry-run shows
the plan without contacting Tinker. Details and memory needs:
references/export-and-serve.md.
Step 9. Convert to GGUF and run with llama.cpp
LLAMA_CPP=~/src/llama.cpp scripts/make_gguf.sh ./export/my-coder/hf-merged ./export/my-coder/gguf my-coder Q4_K_M Q8_0 scripts/serve_gguf.sh ./export/my-coder/gguf/my-coder-Q4_K_M.gguf
make_gguf.sh converts to BF16 GGUF with the checkout's convert_hf_to_gguf.py,
quantizes with the llama-quantize built from the same commit, exports the vision
projector when the model has one, and optionally the MTP draft head (MTP=1).
serve_gguf.sh starts llama-server with the settings that match training:
--jinja, --reasoning-preserve, the template kwarg for preserved thinking, a
context of at least the training window, and the base model's sampling defaults
(temperature 1.0, top-p 0.95, top-k 20, min-p 0). Smoke-test one prompt with a
tool before trusting it, then benchmark against a same-recipe base GGUF.
Rules that saved us the most time
- Never hardcode a renderer name; resolve it from the model with
model_info.get_recommended_renderer_name. Never reuse a dataset rendered for one model with another model. - Never anonymize with one constant fake world. Randomize per session.
- Never train rows without the system prompt and tool definitions the model will be served with.
- Never trust a loss curve alone. Sample.
- Never serve with a chat template that strips earlier-turn thinking when training kept it, and never with llama.cpp's default min-p: match the template kwarg and the base model's sampling defaults.
- Never put data,
.env, run or export directories in a public repo: the pack's.gitignorecoversdata/,runs/,export/,wandb/,.env.
Adapting to other agents
The normalizer reads Claude Code's JSONL. For Codex, OpenCode or your own harness,
write a normalizer that emits the same per-session file (references/claude-code-transcripts.md
documents the message schema); every later step is format-independent.


