Ganesh Mini
Ganesh Mini is a text-only coding fine-tune of
Qwen/Qwen3.5-2B, pinned to base
revision 15852e8c16360a2fea060d615a32b45270f8a8fc. It was trained
jointly for Go, Java, JavaScript, TypeScript, Python, and Rust. The intended
serving mode is non-thinking text generation with an 8,192-token context.
Vision behavior has not been evaluated.
Files
- Repository root: merged BF16 Transformers checkpoint and tokenizer.
gguf/ganesh-mini-1.0-f16.gguf: F16 text target.gguf/ganesh-mini-1.0-q8_0.gguf: Q8_0 text target.adapter/: original rank-8 LoRA adapter. Its configuration retains the training container's/modelbase path; for PEFT use, load the exact Qwen base revision explicitly and then apply this adapter.
The GGUF target exports exclude the unused speculative MTP head. Both GGUF files and the merged Transformers checkpoint loaded and generated a short text-only Python-function smoke response locally. This is an export smoke, not multi-language equivalence or laptop RAM certification.
Public assessment
We ran a local, paired 30-task-per-language public diagnostic using pinned MultiPL-E HumanEval translations for Go, Java, JavaScript, TypeScript and Rust, and the pinned HumanEvalPlus file from the MultiPL-E repository for Python. Both models used Q8_0 GGUF, the same non-thinking text-only serving settings, and the published benchmark verifier in a network-disabled container.
| Language | Ganesh Mini | Qwen upstream | Paired wins / losses |
|---|---|---|---|
| Go | 9/30 | 8/30 | 1 / 0 |
| Java | 12/30 | 13/30 | 0 / 1 |
| JavaScript | 12/30 | 12/30 | 0 / 0 |
| TypeScript | 13/30 | 14/30 | 0 / 1 |
| Python | 18/30 | 18/30 | 0 / 0 |
| Rust | 6/30 | 6/30 | 1 / 1 |
| Total | 70/180 | 71/180 | 2 / 3 |
The fixed sample, exact revisions, hashes, settings, sandbox limits and Rust stop-token correction are documented in the public benchmark method. The paired result report contains task IDs, statuses and evidence hashes, but no benchmark prompts, tests or raw model completions. The initial Rust assembly error is retained locally and excluded; the reported Rust result re-executes the same saved outputs after applying the published stop token to both models.
This is a small locally run public slice, not a full official benchmark leaderboard result or evidence of general improvement. Public tasks may have appeared in upstream pretraining. No private evaluation scores are published.
Full public task diagnostic
We subsequently ran the full pinned public task selection with the same candidate and upstream Q8_0 serving contract and the network-disabled verifier. This run includes all 952 tasks across the six lanes; it does not include the private development or decision sets.
| Language | Ganesh Mini | Qwen upstream | Paired wins / losses |
|---|---|---|---|
| Go | 45/154 | 43/154 | 2 / 0 |
| Java | 51/158 | 52/158 | 0 / 1 |
| JavaScript | 64/161 | 64/161 | 1 / 1 |
| TypeScript | 59/159 | 59/159 | 2 / 2 |
| Python | 84/164 | 86/164 | 0 / 2 |
| Rust | 35/156 | 35/156 | 3 / 3 |
| Total | 338/952 | 339/952 | 8 / 9 |
The full-run method records the pinned inputs, postprocessing, and verifier limits. The sealed paired report records per-task statuses and evidence hashes without prompts, tests, raw outputs, or private assessment. The result is one pass below upstream overall; it does not support a general improvement claim. This is a local public diagnostic, not an official leaderboard submission. The earlier 30-task sample is a subset of these public tasks and must not be added to this total.
Training and limitations
The selected checkpoint completed one unique epoch over an admitted 900-row
package (720 capability examples and 180 broad replay examples), with
651,639 input tokens and 113 optimizer steps. Training used BF16
attention-only rank-8 LoRA, response-only loss, and a 2,944-token training
sequence limit; the selected learning rate was 1e-5. Adapter save/reload
equivalence passed. These figures describe the completed training run, not a
claim of coding superiority.
The 8 GB Q8 and 16 GB F16 laptop profiles have not been certified. Neither the GGUF files nor the merged BF16 checkpoint has completed the full six-lane export-equivalence evaluation. The model may produce incorrect, unsafe, or non-compiling code; execute its output only in a resource-limited, network-disabled sandbox. The original Qwen model license is Apache-2.0. No training data or private evaluation prompts/tests are included here.
Reproducibility identifiers
- Training run evidence seal:
21d78a7e339fb75dd89bdd9ada83a53bdbb10678bd7c3c9b09d182786b698537 - Adapter safetensors SHA-256:
03f8d725c9cad5ce63e524db45883eceb2ab610cf207e6153b004215ec798896 - GGUF F16 SHA-256:
03e6c13f3e75dc322a207d9f6dd96a3c97adcefcaea7f204e77cf8f8fb761d47 - GGUF Q8_0 SHA-256:
f4a3a394e4551211b4de0a58ad5cea1382e5b88ff450c996f1452d3c22388d89
These identifiers permit audit against the retained project evidence without publishing private tasks or raw development outputs.