llama31-8b-knapsack-lora-stateless
Anonymous supplementary release for a double-blind workshop submission. This is one of four LoRA adapters (Mistral-7B-v0.3 / Llama-3.1-8B base model x persistent/stateless training regime) fine-tuned on the Opaque Knapsack agentic task, extending a prior single-base-model result (see the sibling Qwen3-8B release) to a second base model family for the same reproducibility review.
- Base model: meta-llama/Llama-3.1-8B
- Training regime: stateless (trained with a stateless Python interpreter runtime (interpreter state is reset every agent turn))
- Seed: 3407
Training configuration
Fine-tuned with Axolotl, LoRA adapter, 4-bit NF4 quantized base:
| Hyperparameter | Value |
|---|---|
| lora_r | 64 |
| lora_alpha | 128 |
| lora_dropout | 0.05 |
| lora_target_modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| learning_rate | 1e-4 |
| lr_scheduler | cosine |
| optimizer | adamw_torch |
| epochs | 3.0 |
| micro_batch_size | 1 |
| gradient_accumulation_steps | 16 |
| sequence_len | 16384 |
| sample_packing | false |
| seed | 3407 |
| training data | paired traces for the "stateless" regime (see paper Appendix for pairing/filtering procedure) |
Base checkpoint's own chat-format tokens (e.g. <|eot_id|>) are untrained on this base (non-instruct) checkpoint -- trained and served with a hand-written minimal template using only real trained tokens (BOS/EOS + plain-text role prefixes), not Llama's native instruct template.
Provenance
Released anonymously alongside a NeurIPS workshop submission for reproducibility review. Non-anonymous release (paper citation, full code, full training traces) will follow after the review process concludes.