Status
- Base model:
Qwen/Qwen3-8B
- Adapter type: PEFT LoRA
- Status:
stable_seed_adapter
- Intended next stage: staged 1M coding dataset training
Benchmark Summary
Table with columns: Metric, Value| Metric | Value |
|---|
| Benchmark version | fast10_corrected_keep_tokens_standalone_v1 |
| Benchmark count | 10 |
| Max new tokens | 800 |
| Candidate adapter | failed2_exact_v2_dedup_safe |
| Candidate average score | 99.2 |
| Pass 70+ | 10/10 |
| Strong 85+ | 10/10 |
| Fatal regressions | 0 |
| Average latency sec | 19.12 |
| Decision | CANDIDATE_CAN_BE_PROMOTED_AFTER_MANUAL_REVIEW |
Why this adapter exists
The purpose of this stage was to create a reliable stable seed adapter before starting the larger 1M training run.
The seed adapter passed:
- LoRA artifact validation
- Smoke testing
- Corrected benchmark comparison
- Regression checks for Korean honorific style
- Checks for think-tag leakage
- Checks for Chinese character leakage
- Repetitive collapse checks
- MLflow/MLF promotion tracking
Included files
adapter_config.jsonadapter_model.safetensors- tokenizer files, if available
metadata.jsonreports/benchmark_summary_public.jsonreports/benchmark_scores.csv, if available
Loading example
Install dependencies first:
pip install transformers peft accelerate safetensors
Example:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_model = "Qwen/Qwen3-8B"
adapter_id = "YOUR_HF_USERNAME/qwen3-8b-coding-stable-seed-adapter"
tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
base_model,
torch_dtype=torch.bfloat16,
trust_remote_code=True,
device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()
Planned next stage
- Full20 corrected benchmark
- 1M dataset audit
- 1k canary training
- 10k staged training
- 100k staged training
- 1M full training
Notes
This adapter should be treated as a validated seed checkpoint for continued training, not as the final production 1M model.