BoostedV1-ILv9
Iterative-Learning (IL v9) version of BoostedV1: continued LoRA training of deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B. Built from the BoostedV1 Phase-1 model (400 MLX steps + 550 continuation steps) plus 3 rounds of IL v9 self-improvement (generate -> verify -> self-audit -> train on filtered data).
Training
- Base: BoostedV1 Phase-1 (DeepSeek-R1-Distill-Qwen-1.5B + LoRA r=8)
- IL v9: 3 rounds, each generating 600 candidate solutions, externally verifying, model self-auditing (threshold 5/10), then training 100 steps on 52-65 high-quality examples per round
- Best round: round_001 (merged here)
Evaluation
| Benchmark | BoostedV1 (phase-1) | BoostedV1-ILv9 |
|---|---|---|
| GSM8K | 46.0% | 46.5% |
| HumanEval (pass@1) | 7.3% | 11.0% |
The IL v9 loop produced the largest gains on code generation (HumanEval pass@1 7.3% -> 11.0%).
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained('auryn-macmillan/boostedv1-ilv9') tok = AutoTokenizer.from_pretrained('auryn-macmillan/boostedv1-ilv9') inputs = tok('Write Python to check if a number is prime.', return_tensors='pt') out = model.generate(**inputs, max_new_tokens=256) print(tok.decode(out[0]))
Notes
- Standard Qwen2 architecture, no custom code, no trust_remote_code needed.
- Trained in an isolated container; repo contains no training code or data.