hex-lora-opus-magnum-full-solve-results

Qwen3.5-27B (base + DSL LoRA + Python LoRA) evaluated on the full-solve variant of the Opus Magnum eval set: a 29-puzzle holdout where the agent receives only the puzzle (no mechanism / arms / placement scaffolding) and must produce a complete .solution file accepted by the in-game verifier.

This is the hardest of the four difficulty rungs in our variant ladder (instructions-only → mechanism-given → arms-tracks-given → placement-given → full) and the sister of hex-lora-opus-magnum-instructions-only-results-full.

The agent runs an interactive Python REPL and must submit() a working .solution file to the in-game verifier.

Files

• samples.jsonl — 522 rows, one per trajectory across all 3 variants.
• 3 .eval files — one per variant (inspect-ai log format). Open with inspect view <file> or inspect_ai.log.read_eval_log(...).

Top-line accuracy (k=6, partial-credit reward)

Sorted high → low by pass rate. Wilson 95% CI on the binomial proportion shown for context (treats trials as i.i.d. — true uncertainty is wider since trials within a task are correlated).

variant	k	solved/n	pass	95% CI	mean_r
27b-lora-python	6	44/174	25.3%	19.4 – 32.2	0.212
27b-lora-dsl	6	28/174	16.1%	11.4 – 22.3	0.138
27b-base	6	23/174	13.2%	9.0 – 19.1	0.118

pass = fraction of attempts that reached solved=True from the in-game verifier. mean_r is mean of best_reward per attempt under the partial-credit reward profile (cost-based score for solved attempts; ≤ 0.5 partial credit for unsolved-but-progressed attempts).

Variant provenance

variant	run_dir	k	max_tokens
27b-base	opus-27b-base-full29-k6	6	4096
27b-lora-python	opus-27b-py-step210-full29-k6	6	4096
27b-lora-dsl	opus-27b-dsl-step210-full29-k6	6	4096

LoRA Tinker URIs:

• python: tinker://c533f7ee-d0f6-567f-90ec-90ecc549733c:train:0/sampler_weights/000210
• dsl: tinker://6b93f86a-7788-58f9-ac5a-4a33711d8367:train:0/sampler_weights/000210

Eval task

29-puzzle full-solve set (opus-magnum-sean-modal-full-29) split as: 9 campaign puzzles, 6 eval-fixed generated holdout puzzles, 14 minimal handcrafted micro-puzzles. Each task is presented as the raw puzzle with no mechanism / placement / track hints — agents must invent the solver from scratch.

Agent uses the in-house REPL harness (agents/om_repl_runner.py) with:

param	value
max_turns	30
max_tokens	4096
temperature	1.0 (Tinker)
n_attempts	6
scoring_profile	partial-credit
enable_thinking	off

Schema (samples.jsonl)

field	type	notes
variant	str	one of the 3 variant names above
model_id	str	Qwen/Qwen3.5-27B
model_size	str	27B
is_lora	bool	True for LoRA variants
lora_path	str|null	Tinker sampler-checkpoint URI for LoRAs
base_path	str|null	Sampling endpoint for base variant
representation	str	python / dsl / n/a
task	str	full task dir name, e.g. campaign_full__p010_refined_gold__sum
task_bucket	str	campaign / eval_fixed / minimal
sample_id	str	task name with bucket prefix stripped
epoch_k	int	attempt index in 0..5
solved	bool	scorer verdict from in-game verifier
best_reward	float	partial-credit reward in [0, 1]
num_agent_turns	int	assistant messages used
input_tokens, output_tokens	int	aggregated over the trial
final_assistant_text	str	empty here — full transcripts live in .eval files

Loading

import polars as pl
df = pl.read_ndjson("samples.jsonl")
print(
    df.group_by("variant")
      .agg(pl.col("solved").mean().alias("pass"),
           pl.col("solved").sum().alias("n_solved"),
           pl.len().alias("n"))
      .sort("pass", descending=True)
)

To replay a trajectory with the inspect viewer:

pip install inspect-ai
inspect view 27b-lora-python.eval

Provenance

Generated by scripts/run_local_repl_eval.py in opus-magnum-rle on 2026-05-04. Per-run summaries and trajectory JSONs live under jobs/local/<run-id>/ in that repo. Inspect-ai logs combined via scripts/build_full_solve_results_dir.py.

License

MIT. Underlying Qwen base-model license applies to model outputs.