connections-rl-grpo-7b-seed2 (7B)

GRPO (verifiable-reward RL) QLoRA adapter for Qwen/Qwen2.5-7B-Instruct, warm-started from connections-rl-sft-7b and trained with seed 2.

Replication seed 2 of the 7B negative result. Like seed 0, this run collapses semantic grouping below the untrained base model (1.7% of groups against base 4.0%) while driving invalid outputs to 1.2%.

Reading the numbers. Groups correct is a mean count on a 0-4 scale: each board has 4 groups, so the SFT value of 0.346 below means 0.346 of 4 groups per board, i.e. 8.6% of groups. The % of groups column is that value divided by 4. Invalid rate is shown as a percentage in the arm-comparison table and as a bare 0-1 fraction in the seed table, matching how each is stored; both rows are labeled with their units. Values quoted from results-analysis/ (pass@k, entropy/KL) are already 0-1 fractions.

Model Details

• Developed by: Jackson Lukas
• Model type: QLoRA adapter (rank 16, alpha 32, all-linear) for a decoder-only causal LM
• Base model: Qwen/Qwen2.5-7B-Instruct
• Training stage: GRPO (verifiable-reward RL) on top of the SFT warm start
• Language: English
• License: MIT (the base model carries its own Qwen license)

Model Sources

• Repository: https://github.com/jacksonmlukas/connections-rl
• Technical report: report/findings.md
• Full result tables: report/results.md
• Implementation notes: report/implementation_notes.md
• Raw eval artifacts: jacksonlukas/connections-rl-results
• This adapter: jacksonlukas/connections-rl-grpo-7b-seed2

Part of connections-rl. This adapter exists to answer the obvious objection to a single RL run. It re-runs GRPO from the same SFT warm start with a different seed, so that only RL run-to-run variance differs. The primary adapter is connections-rl-grpo-7b; across-seed statistics are in results-seeds/.

Intended Uses

Direct Use

Variance evidence. Use it together with seeds 0 and 1 to check that the reported effect is not a single unlucky draw.

Downstream Use

The adapter and the surrounding harness are intended as a reproducible artifact for research on reward design and reward over-optimization in verifiable-reward RL. The reward, the leakage-aware splits, the evaluation harness with bootstrap CIs and paired tests, and every checkpoint are public so the finding can be re-derived or contradicted.

Out-of-Scope Use

This is a research artifact, not a product. Do not use it to:

• solve live NYT Connections puzzles competitively or to build a puzzle-solving service; held-out solve rate is at or near zero for every arm in this study;
• draw conclusions about the base model's general capability, since these adapters are narrowly specialized and at least one of them measurably degrades the base model's ability on this task;
• perform any high-stakes reasoning task. Nothing here was evaluated for safety, toxicity, factuality or robustness outside the Connections task.

Bias, Risks, and Limitations

Same limitations as the primary adapter: held-out solve rate is 0 of 162, and this adapter is worse than the untrained base at grouping (1.7% of groups against base 4.0%). This replicate was evaluated in session B only, alongside the other seeds.

Study-level limitations that apply to every adapter here:

• One task (NYT Connections) and one reward design. The conclusion is about this reward's structure/semantics decomposition, not about GRPO in general.
• 807 training boards is small for RL. Memorization is a plausible consequence of data scale as much as of the algorithm.
• Checkpoint-level analyses (entropy, KL, phase transition) come from seed 0 at 7B only, because only that run was Hub-synced during training. Endpoint claims carry n=3 seeds per scale; the timing of the collapse carries n=1.
• NYT Connections boards encode US-centric cultural and idiomatic knowledge, so performance is not representative of word-association ability in general.

Recommendations

Cite the across-seed mean and standard deviation (grouping 0.045 ± 0.022 on the 0-4 scale) rather than any single seed.

How to Get Started

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16,
)
base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct", quantization_config=bnb, device_map="auto"
)
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "jacksonlukas/connections-rl-grpo-7b-seed2")
model.eval()

words = ["HAIL", "RAIN", "SLEET", "SNOW", "BUCKS", "HEAT", "JAZZ", "NETS",
         "OPTION", "RETURN", "SHIFT", "TAB", "KAYAK", "LEVEL", "MOM", "RACECAR"]
prompt = tok.apply_chat_template(
    [{"role": "user", "content": "Group these 16 words into 4 groups of 4:\n"
       + ", ".join(words)}],
    tokenize=False, add_generation_prompt=True,
)
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device),
                     max_new_tokens=256, do_sample=False)
print(tok.decode(out[0], skip_special_tokens=True))

The exact prompt template, parser and reward used in the paper are in src/connections_rl. Serving the adapter behind vLLM (--enable-lora) is supported via docker compose up in the repo.

Training Details

Training Data

1,078 NYT Connections puzzles from the gvc-local tagged puzzle database, split strictly chronologically so that every evaluation puzzle postdates every training puzzle:

Split	n	Date range
train	807	2023-06-12 to 2025-08-27
val	108	2025-08-28 to 2025-12-14
test	162	2025-12-15 to 2026-05-29

The chronological split is the leakage control: NYT Connections boards are published daily and widely discussed online, so a random split would let a model benefit from puzzles whose answers circulated before its pre-training cutoff.

Reward Function

Deterministic and unit-tested, with no learned reward model:

Component	Value
Format validity (all 16 board words, 4x4, each used once)	+0.1
Correct groups	+1.0 x (correct / 4)
Full solve bonus	+0.5
One-away shaping	+0.05
Invalid output penalty	-0.1
Maximum	1.6

Only the structural component is cheaply verifiable per sample. The study's central finding is that GRPO optimizes that component at the expense of the semantic component, which is the part the reward cannot check directly.

Training Procedure

TRL GRPOTrainer over the SFT warm start. K=8 completions per puzzle, group-relative advantage, KL penalty to the frozen SFT reference.

Hyperparameter	Value
num_generations (K)	8
temperature	0.9
beta (KL to SFT reference)	0.001
learning_rate	5e-6
per_device_train_batch_size	2
gradient_accumulation_steps	8
num_train_epochs	1
max_completion_length	512

scale_rewards, loss_type, num_iterations and the clipping epsilons were not set explicitly and therefore inherited TRL defaults (scale_rewards="group", loss_type="dapo"). This is recorded deliberately, because scale_rewards="group" is the normalization Dr. GRPO (2503.20783) argues against. See implementation_notes.md.

Why 1 epoch here and 2 at 1.5B. This is a compute-budget constraint, not a tuning choice: at roughly 2-3x the 1.5B step time, a second 7B epoch would have exceeded Kaggle's 12-hour batch limit. It is recorded in configs/train/grpo-7b.yaml. The 7B run still reached its collapsed fixed point well inside one epoch: 98.7% of the total KL displacement was spent by step 150 of 403, so the shorter schedule does not explain the outcome.

Identical to seed 0 except for --seed 2 and the output paths, via the seed-replicate CLI overrides in train/grpo.py.

Evaluation

Testing Data, Factors and Metrics

Held-out test split of 162 puzzles (2025-12-15 to 2026-05-29), strictly after every training date. Greedy decoding via vLLM. Bracketed values are percentile bootstrap 95% CIs over 1,000 resamples. Comparisons between arms use exact McNemar tests on solve rate and paired bootstrap on per-puzzle reward. Results are stratified by puzzle category (wordplay, cultural, category, tag-fillin, silent-letter) in the underlying JSON.

Numbers below come from session B (seed-replication run). See Reproducibility.

Results

This adapter (seed 2), measured alongside its own SFT baseline in the same session:

Metric	SFT baseline	this adapter
Solve rate	1.2%	0.0%
Groups correct (0-4)	0.321	0.068
Groups correct (% of groups)	8.0%	1.7%
Invalid outputs (%)	22.8%	1.2%
Mean reward	0.189	0.141

The SFT baseline above is the session B measurement (groups correct 0.321). The primary adapter cards report 0.346 for the same adapter, measured in session A. Both are correct: greedy decoding is not bitwise deterministic across vLLM layouts. See Reproducibility below.

All three 7B GRPO seeds:

Metric	seed 0	seed 1	seed 2	mean ± sd
Groups correct (0-4)	0.025	0.043	0.068	0.045 ± 0.022
Invalid rate (0-1 fraction)	0.006	0.019	0.012	0.012 ± 0.006
Mean reward	0.125	0.129	0.141	0.132 ± 0.008

Reproducibility and Measurement Noise

Some arms were measured in two vLLM serving configurations, referred to in the repository as session A (results/, results-7b/) and session B (results-seeds-*/, aggregated in results-seeds/seed_summary.json).

Greedy decoding is not bitwise deterministic across vLLM batching and parallelism layouts, so a small number of borderline tokens flip. The 7B SFT arm differs on 2 of 162 puzzles for grouping and 3 of 162 for format validity between sessions; every GRPO arm reproduces exactly (0 of 162 on all metrics).

That asymmetry is itself evidence for the entropy-collapse account: the final 7B GRPO policy has a measured entropy of 0.0099 nats/token and therefore has no borderline decisions left to flip, whereas the SFT policy has the highest entropy of any arm (0.303 nats/token). Statistics are never mixed across sessions within a single comparison.

Environmental Impact

Trained on free-tier NVIDIA T4 GPUs (16 GB, Turing, pre-Ampere) via Kaggle and Google Colab. Total training compute for this adapter was approximately about 5.5 hours on a single T4. The entire 7B study, including all seed replicates and evaluation, was run at $0 marginal cost on free-tier hardware. Carbon emissions were not directly measured; the T4 has a 70 W TDP, which bounds the energy use of a single run well below 1 kWh.

Citation

If you use this adapter or the accompanying analysis, please cite the repository:

@software{lukas_connections_rl_2026,
  author = {Lukas, Jackson},
  title  = {connections-rl: What Verifiable-Reward RL Actually Transfers},
  year   = {2026},
  url    = {https://github.com/jacksonmlukas/connections-rl},
  note   = {Two-scale, three-seed GRPO study on NYT Connections}
}

The predecessor multi-agent work is published as Snap Out of It (ACL 2025, REALM Workshop).

Model Card Contact

Open an issue at https://github.com/jacksonmlukas/connections-rl/issues.