Gemma-4-E4B-Sol-Traces-v4
Targeted routing-repair continuation from v3. Continuation-trained on 125 mined trajectories (SWE-bench, synthetic, bash history, GitHub commits) to fix run_command, search_code, and apply_patch tool selection.
Training
| Parameter | Value |
|---|---|
| Base model | unsloth/gemma-4-E4B-it |
| Adapter start | v3 from-scratch (336 LoRA keys) |
| Training type | Continuation (100 steps) |
| Dataset | 125 train / 15 val / 17 test |
| Sources | SWE-bench (50), synthetic (50), bash history (55), GitHub (2) |
| Tool schema | 6 tools (search_code, run_command, apply_patch, read_file, list_files, write_file) |
| Learning rate | 3e-5, cosine, 3% warmup |
| Batch | 1 × 8 grad accum |
| Runtime | 12 min on H100 |
| Training loss | 0.093 |
| Eval loss | 0.671 |
| Cost | ~$0.80 |
Routing Improvement
| Tool | v1 | v2 | v3 | v4 |
|---|---|---|---|---|
| run_command | 2/5 | 0/5 | 0/5 | 5/5 ✅ |
| search_code | 0/5 | 0/5 | 0/5 | 1/5 |
| apply_patch | 1/5 | 0/5 | 0/5 | 0/5 |
| list_files | 5/5 | 5/5 | 5/5 | 4/5 |
| read_file | 4/5 | 5/5 | 5/5 | 5/5 |
| no-tool | 4/5 | 5/5 | 5/5 | 3/5 |
| Overall | 53% | 50% | 50% | 60% |
Files
| File | Size |
|---|---|
gemma-4-e4b-sol-traces-v4-Q4_K_M.gguf |
4.97 GiB |
gemma-4-e4b-sol-traces-v4-f16.gguf |
14.02 GiB |
adapter/adapter_model.safetensors |
35 MiB |
training_stats.json |
— |
Usage
llama-cli -m gemma-4-e4b-sol-traces-v4-Q4_K_M.gguf -ngl 99