DCAgent2/g1_gptlong_plus_diverse_tezos_glm47_traces

132,259 rows. Concatenation of:

• DCAgent/g1_min_episodes_e1_gpt_long_top8_glm47_traces (37,925 rows) — top8 + extra long swegym traces
• DCAgent/g1_diverse_tezos_top4_100k_glm47_traces (94,334 rows) — balanced 4-way mix of swesmith/issue/superuser/tezos

No deduplication; shuffled with seed=42. Schema: intersection columns (agent, conversations, date, episode, model, model_provider, result, run_id, task, trial_name).

Subset breakdown

subset	rows	%
swesmith	27,500	20.8%
issue	27,500	20.8%
tezos	25,485	19.3%
superuser	23,962	18.1%
swegym	16,200	12.3%
scaffold	4,591	3.5%
tor	3,773	2.9%
r2egym	3,248	2.5%

Source models

• gptlong → trained DCAgent/g1_gptlong_top8_32b (Qwen3-32B, lr=4e-5, wd=0.04, gn=1e-3, 5 epochs)
• diverse_tezos → trained DCAgent/g1_diverse_tezos_100k_32b

Generated for combining the two recipes' data.

Training config

The recommended SFT recipe for this dataset is the gptlong/opt100k config — included in this repo at configs/32k_base_bs96_opt100k.yaml.

Key hyperparameters:

	value
base model	Qwen3-32B (override model_name_or_path from the YAML's Qwen3-8B default)
learning_rate	4e-5
weight_decay	0.04
max_grad_norm	1e-3
adam_beta2	0.98
num_train_epochs	5
global_batch_size	96
cutoff_len	32768
template	qwen3 (thinking)
lr_scheduler	cosine, warmup_ratio=0.1
disable_shuffling	true
precision	bf16
deepspeed	ZeRO-3 (ds_z3_accelerate.json)

Suggested cluster: 24 nodes × 4 GPUs = 96 GPUs, gradient_accumulation_steps=1.