SFT tokenization and loss

Kiln native SFT has one Qwen3.5-4B tokenization and label contract. Inline examples, server-local JSONL, and named datasets all reach this same path after message-schema validation.

This token/label contract is one part of the fixed native_online_lora_v1 update contract.

Row admission is fail-closed by default and uses the same tokenization check through every transport. The optional explicit skip policy and stable kept/rejected row hashes are specified in SFT Ingestion, Invalid Rows, and Row Identity.

Conversation rendering

SFT validates the source-pinned model chat_template.jinja, selects TRL 1.8's prefix-preserving Qwen3.5 thinking training variant, and renders with add_generation_prompt=false. A training sequence ends after the final serialized message. Kiln does not append the open assistant/thinking prefix used to start inference.

OpenAI-shaped assistant tool calls are accepted with function.arguments as a JSON string. Kiln parses that string into the object the Qwen template expects before rendering. A tool-call-only assistant message may use content: null or omit content; both normalize to empty content. Tool responses retain role: "tool", name, and tool_call_id through ingestion, although the current Qwen template does not serialize the latter two fields.

Sequence length and truncation

Kiln tokenizes the complete rendered conversation. Native SFT does not silently truncate a row, join multiple rows into one sequence, or split one row across optimizer steps. Each accepted row remains one conversation and one optimizer step.

Ingestion records the actual longest accepted sequence. Server preflight uses that length for memory admission, and the trainer uses each row's actual length when choosing its gradient-checkpoint plan. A row that cannot pass tokenization and label validation is rejected according to the configured invalid-row policy; it is never shortened into a different training example.

Assistant-only labels

The mask contract is kiln.qwen35-assistant-only.v1:

• System, user, and tool-response turns are masked from loss.
• The assistant role header <|im_start|>assistant\n is masked from loss.
• Assistant thinking text, answer text, and Qwen tool-call XML are supervised.
• The closing <|im_end|> and its trailing newline are supervised.
• There is no appended generation prompt in an SFT sequence.

The checked-in fixture uses the standard Hugging Face label representation:

labels[i] = input_ids[i]  when assistant_mask[i] == 1
labels[i] = -100          otherwise

The causal loss predicts token i from the preceding position. Excluding the assistant role header is deliberate and matches TRL; including the message terminator ensures the model is trained to stop. Kiln carries TRL's generation boundaries through rendering as internal markers and removes them before tokenization; it does not infer loss spans by searching message text for Qwen delimiter strings. An internal-marker collision fails the row closed.

HF and TRL equivalence

The official Qwen3.5-4B template at the pinned revision does not contain Hugging Face {% generation %} blocks. Asking Transformers for an assistant mask from that template directly produces an all-zero mask, so it is not a valid assistant_only_loss template by itself.

TRL 1.8 detects this template and selects its bundled qwen3_5_think_training.jinja. That template is prefix-preserving, always serializes the reasoning block on completed assistant turns, starts the {% generation %} span after the assistant role header, and ends it after the message terminator. The oracle runs that exact source-pinned template through Transformers and uses return_assistant_tokens_mask as the label mask. Kiln uses the same template with the two Jinja extension tags adapted to internal span markers that are removed after rendering. The effective-template digest is over the marker-free template. The checked-in real-tokenizer test proves rendered text, token IDs, masks, and labels still match exactly.

The source-pinned fixture covers plain, explicit thinking, tool-call, tool-response, multi-turn, and adversarial delimiter-literal examples. The latter proves message text containing Qwen role and terminator strings cannot move assistant loss boundaries. It records exact rendered text, token IDs, assistant masks, and -100 labels together with the model revision, artifact hashes, Transformers and TRL source hashes, and dependency versions:

• crates/kiln-train/tests/fixtures/qwen35_sft_oracle_v1.json
• scripts/qualification/qwen35_sft_oracle.py

Regenerate and compare it against a staged copy of the pinned model artifacts:

uv run \
  --with transformers==5.13.1 \
  --with tokenizers==0.22.2 \
  --with jinja2==3.1.6 \
  python scripts/qualification/qwen35_sft_oracle.py \
  --model-path /path/to/Qwen3.5-4B \
  --check

KILN_QUALIFICATION_MODEL_PATH=/path/to/Qwen3.5-4B \
  scripts/qualification/cargo-test-bounded.sh test --locked --offline -p kiln-train \
  --test qwen35_sft_oracle -- --ignored

The Python contract tests validate the checked-in fixture without installing Transformers. The ignored Rust integration test is the real-tokenizer gate and must be run on a machine with the exact staged artifacts. Neither test requires a GPU.