Arm C — the all-pass subset of Arm B, used to separate the marginal value of additional data by type. Starting from this same 1,255-trajectory set, adding 785 fully-passing trajectories gives Arm A, while adding 785 imperfect ones gives Arm B, so A−C and B−C are directly comparable increments.

Ablation context

These runs study whether the gain from lowering the trajectory-inclusion threshold comes from data quality or data quantity. Lowering the threshold changes both at once (2,040 -> 3,291 trajectories, +61%), so the arms below hold N fixed and vary only composition.

arm	N	fully-passing	imperfect	imperfect %
A (all-pass baseline)	2,040	2,040	0	0%
B (size-matched mix)	2,040	1,255	785	38%
C (all-pass subset of B)	1,255	1,255	0	0%
D-dose20	2,040	1,632	408	20%
R (full, lowest threshold)	3,291	2,040	1,251	38%

This model: 1,255 trajectories, all fully-passing (0% imperfect).

Training data

Multi-turn bash-agent rollouts. Tool calls are emitted as plain text inside the assistant turn using a <tool_call>{"name": "bash", ...}</tool_call> protocol defined in the system prompt (not native function calling); tool output returns in a tool role. Loss is computed on assistant turns only.

Training procedure

Full-parameter SFT via LLaMA-Factory, DeepSpeed ZeRO-3, BF16, on 8x B200.

hyperparameter	value
learning rate	5e-6
lr scheduler	cosine, warmup ratio 0.1
weight decay	0.0
optimizer	AdamW
epochs	5
per-device batch size	1
gradient accumulation	8
effective batch size	64
max sequence length	20,000
precision	bf16
attention	sdpa

Results

Final train loss 0.3247 (from 0.9182 at the first logged step); mean train loss 0.4421 over 100 steps / 5 epochs. Runtime 1.54h.

Note: a small fraction of samples (~5%) exceed the 20,000-token cutoff and are truncated.

Loss is not comparable across arms: arms containing imperfect trajectories are fitting partly-failing rollouts, so a lower loss does not imply a better agent. These checkpoints must be compared by task success rate, not by training loss.

Framework versions

Transformers 5.8.0
PyTorch 2.11.0+cu130
DeepSpeed 0.18.9
Datasets 4.0.0

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "March07/Qwen2.5-Coder-32B-terminal-agent-sft-allpass1255"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")

messages = [
    {"role": "system", "content": "You are an expert technical assistant with access to bash tools."},
    {"role": "user", "content": "List the files in /workspace."},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))