---

license: apache-2.0

base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct

tags:

• nl2sh
• natural-language-to-shell
• llama-cpp
• gguf
• lora
• sft

datasets:

• westenfelder/NL2SH-ALFA

language:

• en

pipeline_tag: text-generation

---

nl2sh-qwen25-coder-1.5b

Natural-language → shell command generator, drop-in alternative to

whatisit-nl2sh's nl2sh-1.5b

(same base model family: Qwen2.5-Coder-1.5B-Instruct), trained on a broader,

hardened data mix.

**Measured on InterCode-ALFA (300 tasks, temp 0, max_tokens 64, unmodified

upstream scorer with embedding heuristic at threshold 0.75):**

| model | size | pass rate |

|---|---|---|

| GPT-4o (published) | cloud | 0.730 |

| this model, Q4_K_M (941 MB) | 941M | 0.6567 |

| nl2sh-3b (upstream, published) | 1.9G | 0.657 |

| this model, Q4_K_M + imatrix | 941M | 0.6500 |

| this model, Q6_K | 1.2G | 0.6433 |

| this model, Q8_0 | 1.6G | 0.6367 |

| nl2sh-1.5b (upstream, published) | 941M | 0.620 |

| nl2sh-1.5b (re-measured on same rig) | 941M | 0.590 |

Same-size Q4_K_M beats the upstream default by +3.7 points over its published

score and +6.7 points measured head-to-head on identical hardware/protocol.

(Quant-to-quant differences of 1–2 pts are within the ±5.5 pt binomial noise at

n=300 — treat all variants as one ~0.65 blob; Q4_K_M is recommended as it holds

the top as-measured score at the smallest size.)

Files

| file | size | use |

|---|---|---|

| qcoder-nl2sh-q4_k_m.gguf | 941M | default pick — best measured score |

| qcoder-nl2sh-q4_k_m_imx.gguf | 941M | imatrix variant (imatrix.gguf, corpus = NL2SH pairs) |

| qcoder-nl2sh-q6_k.gguf | 1.2G | quality step-up |

| qcoder-nl2sh-q8_0.gguf | 1.6G | near-lossless (+0.14% ppl) |

| qcoder-nl2sh-f16.gguf | 2.9G | reference |

| imatrix.gguf | 1.9M | importance matrix used for _imx build |

| lora-adapter/ | ~180M | PEFT LoRA adapter |

Perplexity ladder (held-out NL2SH pairs): f16 4.649 · Q8 4.655 · Q6 4.665 ·

Q4+imx 4.720 · Q4 4.789.

Robustness

Trained with organic chatter pairs so arbitrary input always yields a sane,

harmless command instead of garbage:

> hello          → echo hello
> thanks         → echo "you're welcome!"
> who am i       → whoami
> qwerty         → pwd
> what time is it→ date

Training setup

| | |

|---|---|

| Base | Qwen2.5-Coder-1.5B-Instruct |

| Data | 120,349 NL/command pairs: NL2SH-ALFA train + tldr-pages + commandlinefu + NL2Bash + cli-1m sample + git-instruction + organic robustness set; length-capped ≤56 tokens, single-line, deduped, decontaminated vs InterCode-ALFA test |

| LoRA rank / alpha / dropout | 32 / 64 / 0.05, all attention+MLP projections |

| LR / schedule | 2e-4, cosine, 3% warmup |

| Epochs / batch / seq | 1 / 32 effective (8×4) / 224 grouped-by-length |

| Loss | assistant-only (chat-template masked) |

| Precision / seed | bf16, seed 42 |

| Hardware | AMD Radeon RX 9070 XT (gfx1201), 2h12m |

Final train loss 0.904 avg, token accuracy 82.7%.

Note for RDNA4/ROCm trainers: requires

PYTORCH_HIP_ALLOC_CONF=roundup_power2_divisions:16 and hipBLASLt disabled

(TORCH_BLAS_PREFER_HIPBLASLT=0 ROCBLAS_USE_HIPBLASLT=0 DISABLE_ADDMM_HIP_LT=1)

to avoid the gfx1201 page-fault bug (ROCm#6600).

Usage

llama-server -m qcoder-nl2sh-q4_k_m.gguf --temp 0

whatisit setup --model ./qcoder-nl2sh-q4_k_m.gguf
whatisit doctor

Greedy decoding, output ≤64 tokens — answers with a single command.

Non-command input still produces a harmless command (see robustness above).

Limitations

• Single-turn, no shell state; English only.
• One command per request, not scripts.
• Quant-to-quant benchmark differences are within measurement noise.