CDBiddulph/wordchain-rl

Wordchain

Word-chain puzzles used to train CDBiddulph/qwen3-30b-a3b-wordchain-rl with reinforcement learning.

Each example asks the model to connect a start word to an end word with the shortest chain of words, where every adjacent pair must appear together in a set phrase (e.g. FAST → FOOD via "fast food"). The model is also asked to critique its own chain.

Fields

• query: the full prompt given to the model
• start_word, end_word: the words the chain must begin and end with

Splits

Split	Examples	Use
train	10,000	RL training prompts
validation	100	Evaluated periodically during RL training
test	100	Held out; not seen during training

How the model was trained

CDBiddulph/qwen3-30b-a3b-wordchain-rl is a LoRA (rank 32) on Qwen/Qwen3-30B-A3B, trained with RL on Tinker using the train split (all 10,000 examples available, seed 1, 4 samples per prompt, batch size 100, LR 1e-4). The released checkpoint is from step 70. The validation split was used for evaluation during training. Prompts were formatted with a DSPy ChatAdapter.

Reward. An LLM judge (gpt-5-mini) reads the response, extracts the chain and the claimed set phrase for each link, and labels each link valid or invalid. The reward is 0 if the chain doesn't start and end with the right words, or if any link is judged invalid or has no stated phrase. Otherwise it is 1.0 - 0.1 * (number of words - 2), so shorter chains score higher. The self-critique the prompt asks for is not rewarded.

Because the reward depends on an imperfect LLM judge, this environment allows reward hacking (e.g. claiming dubious set phrases the judge accepts).

How the data was made

Start and end words were sampled from the 5,000 most frequent English words (wordfreq), filtered to 3–10 letter base-form non-proper-noun words with spaCy. See data_generation/prepare_wordchain.py and the scorer in tasks/wordchain/scorer.py.