davidheineman/tiny-rl-refined-tasks-v5.5

Tiny RL refined tasks v5.5

The default sft and rl_envs configs contain exactly the eight tasks used in the real v5.5 mixture experiment: GSM Partial, MBPP Docstring, Chat, Arithmetic Simple, LaTeX, Brackets, Regex and Chess. Additional tasks remain available through per-task configs and all_sft / all_rl_envs. Exact origins and SHA256 hashes are in dataset.json; checks are in checks/.

Default and per-task SFT configs use plain text, without <|im_start|>, <|im_end|> or <|endoftext|> markers. Legacy turns become Instruction:, Question: and Answer: sections. Token IDs, supervision masks and exact SFT budgets are rebuilt for that serialization. sft_native, all_sft_native and <task>_sft_native preserve the original format used for existing experiments. No models were retrained for this publication.

from datasets import load_dataset
sft = load_dataset("davidheineman/tiny-rl-refined-tasks-v5.5", "sft", split="train")
rl = load_dataset("davidheineman/tiny-rl-refined-tasks-v5.5", "rl_envs", split="train")
# Per-task configs: partial_gsm8k_sft, chess_sft, chess_sft_native, etc.

Task key	Task	Set	Origin	SFT input positions
partial_gsm8k	GSM Partial	v5.5 benchmark	v5.5-primary-eight	2,097,152
mbpp_docstring	MBPP Docstring	v5.5 benchmark	v5.5-primary-eight	2,097,152
chat	Chat	v5.5 benchmark	v5.5-primary-eight	2,097,152
arithmetic_simple	Arithmetic Simple	v5.5 benchmark	v5.5-primary-eight	2,097,152
latex	Valid LaTeX	v5.5 benchmark	v5.5-primary-eight	2,097,152
brackets	Valid Brackets	v5.5 benchmark	v5.5-primary-eight	2,097,152
regex	Valid Regex	v5.5 benchmark	v5.5-primary-eight	2,097,152
chess	Chess	v5.5 benchmark	v5.5-primary-eight	2,097,152
arithmetic	Arithmetic	Additional task	math-full-reference-16m-v1	16,777,216
gsm8k	Hinted GSM8k	Additional task	math-v1-original-full-gsm8k	2,097,152
math	Hinted MATH	Additional task	v3-frozen-predecessor	16,777,216
python_syntax	Valid Python	Additional task	imported-v4-pilot-2m-v2	2,097,152
mbpp	Hinted MBPP	Additional task	code-window4-rl64k-2m-v1	2,097,152
make_sentence	Make sentence	Additional task	imported-v4-pilot-2m-v2	2,097,152
reward_modeling	Reward modeling (Alpaca + Skywork RM)	Additional task	language-original-ppo4-2m-v1	2,097,152
mmlu	Hinted MMLU	Additional task	v3-frozen-predecessor	16,777,216
json	JSON from spec	Additional task	v3-frozen-predecessor	16,777,216
cube	Rubiks Cube	Additional task	v3-frozen-predecessor	16,777,216
arithmetic_simple_shift3	Arithmetic Simple (operand +3)	Scaling variant	v5.5-scaling-operand-plus3	2,097,152

GSM Partial ends with So, the answer is <expression> = <value>. Arithmetic Simple uses equation traces such as Answer: (0 + 8) = 8. The distinct gsm8k and arithmetic configs retain the full reasoning / final Answer: variants as additional tasks. arithmetic_simple_shift3 is the separate operand-plus-3 version used in the scaling sweep.

hf download davidheineman/tiny-rl-refined-tasks-v5.5 --repo-type dataset --local-dir tiny-rl-v55
cd tiny-rl-v55
uv sync
uv run python envs.py task=partial_gsm8k split=rl seed=5917
# For Make sentence:
uv sync --extra language

from envs import Environment
env = Environment(".", "partial_gsm8k", interface="plain")
episode = env.sample("rl", seed=5917)
print(episode.prompt, episode.max_tokens)
print(env.reward(episode, episode.reference))

Use one process per task when tasks have different frozen source versions. The environment defaults to interface="plain"; use interface="native" for original prompt token IDs. Chess uses env.step(episode, uci_move) until done and returns reward 1 only for a complete correct puzzle. Plain formatting preserves the same board positions, candidate lists, moves and grader; it changes only serialization.

Reward modeling uses original Alpaca instructions/answers and pinned Skywork/Skywork-Reward-V2-Llama-3.2-3B revision beafe5f2e6a275ce8c556f68f2d82de7fefbc3d8. env.reward(episode, completion, mode="skywork") returns its raw score. Default mode="study" preserves the current quality-gated study reward, including the external whole-response judge; it needs the inherited OpenAI credentials and downloads the reward model. No reward-model weights are included.

Native rows, tests, hints, demonstrations and source metadata are preserved in row_json and raw/. Token counts and IDs use GPT-2. Legacy SFT documents include authoritative input_ids and loss_mask in document_json, for both formats. Existing native benchmark scores do not claim performance for the new plain serialization. Input positions are distinct from supervised target counts. Run uv run python verify.py to verify the exported data and full-puzzle Chess interface.

The additional gsm8k task restores the original frozen word problems and full multi-step references from refined-math-v1/pilot-2m, before the supplied-final-calculation rewrite. Its questions do not include an added final-calculation hint. It still uses the historical 1–16-token completion interface and canonicalized worked solutions, rather than verbatim original answer prose. This restoration does not change partial_gsm8k or the eight-task benchmark. Hinted MATH, MMLU, JSON and Cube retain their frozen predecessor definitions. MMLU includes supplied oracle solution traces; it is not standard MMLU evaluation. MBPP includes clear implementation rewrites with original tests; it is not a canonical-reference or unhinted benchmark. Hint ranges and decoding temperatures are recorded per task.

Source-group splits are preserved and checked for disjointness within each task. Source licenses continue to apply; this collection does not grant a new common license. The inherited Tectonic Linux x86_64 executable/cache and pinned spaCy English model are included for the local graders. Python environments support CPU and CUDA; external judge/model inference was not rerun as part of this data-only export.