GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
March07
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
March07
Model Tree
Qwen/Qwen2.5-Coder-32B-Instruct
Input Modalities
Output Modalities
Supported Functionality
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: 2,040 trajectories: 1,632 fully-passing + 408 imperfect (20% imperfect).
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.
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 |
Final train loss 0.2765 (from 0.9122 at the first logged step); mean train loss 0.3887 over 160 steps / 5 epochs. Runtime 2.45h.
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.
python
from transformers import AutoModelForCausalLM, AutoTokenizermodel_id = "March07/Qwen2.5-Coder-32B-terminal-agent-sft-mix2040-dose20"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))
| precision | bf16 |
| attention | sdpa |