feat: add --sft-assistant-turns option for multi-turn SFT

[Liu14159]

Summary

Add --sft-assistant-turns [all|last] option for multi-turn SFT, allowing users to train on every assistant turn (default all) or only the final assistant turn (last). Earlier assistant turns in last mode are treated as context. User, system, and tool-result tokens are always excluded from the loss.

Usage

# Train on all assistant turns (default, backward compatible)
areno train \
  --algo sft \
  --ckpt Qwen/Qwen3-0.6B \
  --dataset-path /path/to/multiturn.jsonl \
  --dataset-loader-fn examples/sft/multiturn/dataset_loader.py \
  --sft-assistant-turns all \
  --tp-size 1 \
  --world-size 1 \
  --batch-size 2 \
  --mini-bs 1

# Train only on the final assistant turn
areno train \
  --algo sft \
  --ckpt Qwen/Qwen3-0.6B \
  --dataset-path /path/to/multiturn.jsonl \
  --dataset-loader-fn examples/sft/multiturn/dataset_loader.py \
  --sft-assistant-turns last \
  --tp-size 1 \
  --world-size 1 \
  --batch-size 2 \
  --mini-bs 1

Multi-turn data format (JSONL, one conversation per line):

{"messages": [{"role": "user", "content": "What is 2+2?"}, {"role": "assistant", "content": "4"}, {"role": "user", "content": "And 3+3?"}, {"role": "assistant", "content": "6"}]}

Related issue

Fixes #224

Type of change

• ✨ New feature

What changed

File	Change
areno/api/trainer_config.py	+3: add sft_assistant_turns field with __post_init__ validation
areno/api/data_utils.py	+74: new messages_to_tokens_and_mask() with last_assistant_only support
areno/api/trainers/sft.py	+36/-16: integrate messages format + sft_assistant_turns passthrough into _record_to_train_sequence
areno/cli/train.py	+10: expose --sft-assistant-turns CLI option, config summary, dashboard settings
tests/test_trainer_dataset_utils_cpu.py	+210: 13 new CPU tests (MultiTurnSFTMaskTest)
docs/cli/training.rst	+40: document option + runnable example
docs/cli/dataset_loaders.rst	+18: multi-turn messages format documentation
examples/sft/multiturn/	New: README.md + dataset_loader.py

Edge cases handled

Scenario	Behavior
{"messages": None}	Returns None (filtered out)
{"messages": []}	Returns None (empty, no trainable tokens)
Messages with no assistant turn	Returns None (nothing to train)
Invalid sft_assistant_turns value	Raises ValueError with clear message
Single-turn prompt/response rows	Unchanged — sft_assistant_turns has no effect
Neither messages nor prompt/response	Raises ValueError with schema hint
last mode with single assistant turn	Identical to all mode

Design decisions

• Mask convention: True = do not train (context), False = train (assistant response). Matches existing prompt_response_to_tokens_and_mask contract.
• Default all: preserves backward compatibility — existing single-turn SFT runs are unaffected.
• Chat-template path: encodes conversation incrementally (messages[:i+1]) to attribute tokens to individual turns. O(N) calls to apply_chat_template for N messages; acceptable for typical SFT data (2-10 turns).
• Plain-text fallback: for tokenizers without chat_template, concatenates role: content per turn.
• EOS handling: appended after last message if not already present, marked as trainable in both modes.
• No new dependencies: reuses existing data_utils, tokenizer, and TrainerConfig contracts.

How was it tested?

Tested on Kaggle (Python 3.12, T4 GPU, ARENO_BUILD_EXT=0):

pytest tests/test_trainer_dataset_utils_cpu.py -v

25 tests collected, all passed. Config validation and CLI option verified separately:

TrainerConfig default: sft_assistant_turns='all'
TrainerConfig(sft_assistant_turns='last'): accepted
TrainerConfig(sft_assistant_turns='invalid'): raises ValueError
areno train --help | grep sft-assistant-turns: --sft-assistant-turns [all|last]

Checklist

• The PR title summarizes the contribution.
• Linked the related issue in the description (if any).
• Existing tests pass (pytest tests/ -k cpu).
• New behavior is covered by tests.
• Described the test commands run and any hardware limitations.
• Public API / CLI changes are additive and backward-compatible (see CONTRIBUTING.md).