Qwen3-8B CodeAct Stateless LoRA — Opaque Knapsack (seed 1337)

QLoRA (4-bit, r=64, alpha=128) fine-tune of Qwen/Qwen3-8B on CodeAct agent traces generated under a stateless Python interpreter runtime, trained to solve the Opaque Knapsack task (partially-observable knapsack requiring iterative state revision via a budgeted inspect() tool).

This is one run of a 2-seed variance sweep (seed=1337, plus a sibling seed=777 run) reproducing the original seed=3407 result from the paper "Agents Learn Their Runtime: Interpreter Persistence as Training-Time Semantics" (Victor May, Aaditya Salgarkar, Yishan Wang, Diganta Misra, Huu Nguyen, arXiv:2603.01209). The paper's core finding: a model trained on persistent-runtime traces learns to rely on interpreter state persisting across turns; when forced into a stateless runtime at inference it suffers cascading NameError/recovery loops (the "amnesia tax"). See the GitHub repo for the full pipeline.

Training data

Source: pythonformer/agents-learn-runtime-train, data/knapsack/stateless/traces.jsonl — 1000 CodeAct agent traces generated under the stateless runtime, teacher model Gemini (see paper Table 1, 17.8-22.8% teacher exact-solve rate).
Copied byte-identical (sha256-verified) into the training pipeline, zero transformation.
Not a re-derived/reconstructed trace set — this repo's earlier attempts to regenerate an equivalent set locally (prepare_paired_data.py, 1109 pairs) were found to NOT match this canonical data and were discarded; see HANDOFF_knapsack_seed_sweep.md in the repo for the full story.

Hyperparameters


Base model	`Qwen/Qwen3-8B`
Quantization	4-bit QLoRA (bitsandbytes NF4)
LoRA rank / alpha / dropout	64 / 128 / 0.05
LoRA target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training context	16384 tokens
Batch size	1 (micro) x 16 (grad accum) = 16 effective
Epochs	3 (~45 optimizer steps over 1000 examples)
Optimizer / LR / schedule	AdamW (torch) / 1e-4 / cosine, 3% warmup
Training seed	1337
Thinking	disabled (`enable_thinking: false`, i.e. `/nothink`)

Final training loss: 0.1313 (this run) — manifest and full Axolotl config recorded in manifest.json / axolotl_config.yaml in this repo, also mirrored into this HF repo for provenance.

Benchmark results (this checkpoint: stateless-trained, seed 1337)

n=300 tasks/cell, from pythonformer/agents-learn-runtime-tasks (easy/medium knapsack splits). "matched" = eval runtime equals the training runtime; "cross-eval" = eval runtime differs from training runtime (tests the amnesia-tax hypothesis). Metric is score (metrics.score.mean, partial-credit normalized optimality = agent's basket value / optimal basket value, 0-1) — used instead of the stricter exact-optimum-match rate (average_is_solved), which is too sparse at n=300 for this task to compare cleanly. Three training seeds shown side by side, no averaging — seed 3407 is the initial run (source: pythonformer/agents-learn-runtime-benchmarks, gated dataset, request access on the dataset page, n=300/cell), seed 1337 / seed 777 are this variance sweep.

eval runtime	difficulty	seed 3407 (initial)	seed 1337	seed 777
persistent (cross-eval)	easy	0.294	0.143	0.155
persistent (cross-eval)	medium	0.092	0.015	0.032
stateless (matched)	easy	0.614	0.137	0.117
stateless (matched)	medium	0.331	0.014	0.002

Seed 3407 numbers use the dataset's folder naming {runtime}_{model}_knapsack_{difficulty} — confirmed by downloading a cell's config.yaml and reading its model:/persistent_state: fields directly (the dataset's own README states the opposite order, {model}_{runtime}, but that does not match this snapshot's actual folder structure).

Base model reference (seed-independent, not this checkpoint)

eval runtime	difficulty	seed 3407 (initial)	seed 1337 / seed 777 (base, seed-independent)
persistent	easy	0.050	0.061
persistent	medium	0.030	0.030
stateless	easy	0.034	0.035
stateless	medium	0.030	0.017

Related checkpoints

Citation

@article{pythonformer260301209,
  title   = {Agents Learn Their Runtime: Interpreter Persistence as Training-Time Semantics},
  author  = {Victor May, Aaditya Salgarkar, Yishan Wang, Diganta Misra, Huu Nguyen},
  journal = {arXiv preprint arXiv:2603.01209},
  year    = {2026}
}