Qwen3-0.6B · F5 Terminal Balanced
F5 freezes the Unsloth UD-IQ2_M base. For each of 196 projections, a rank-eight correction is reconstructed from packed INT4 codes by a shared 18→64→64→8 decoder with layer-specific FiLM gains and biases. We trained the codes and all 13,064 decoder/FiLM parameters together, giving terminal turns extra weight (1.75; first-token factor 2.5) and using a language KL term (0.05). The base plus compact correction occupied 271.35 MB in the original experiment. The exported adapter is 10.12 MB; it must be loaded alongside the 268.70 MB base at scale 1.
Project repository: compression code and experiments · Model recipe.
Upstream: Qwen/Qwen3-0.6B, pinned checkpoint. Method, code and reproduction files · Machine-readable scores · SHA-256 checksums.
Run locally
Download the files you need with the Hugging Face CLI:
hf download rexarrior/Qwen3-0.6B-F5-Terminal-Balanced --local-dir F5-Terminal-Balanced
cd F5-Terminal-Balanced
llama-server -m Qwen3-0.6B-UD-IQ2_M.gguf --lora Qwen3-0.6B-OpenResearch-F5-Terminal-Balanced-Full-Decoder.gguf -c 2048
Inference weights on disk: 278.82 MB (decimal; excludes the engine).
In-house evaluation
Counts are passes out of all tasks. Fresh, A6 and A8 score complete trajectories; A6 and A8 are known task sets. The language perplexity uses the same 2,040 scored tokens from Alice in Wonderland on CPU; lower is better.
| Check | Result |
|---|---|
| Fresh tool-use trajectories | 5/48 |
| Known A6 trajectories | 12/48 |
| Known A8 trajectories | 9/48 |
| Next action from supplied history | 8/24 |
| Full continuation from supplied history | 0/24 |
| Strict English | 20/96 |
| Alice perplexity ↓ | 38.7522 |
External evaluations
BFCL v4: 21 categories without web search. Family accuracies are shown separately; these are not a full leaderboard overall. API-Bank uses 744 offline-evaluable calls; IFEval scores prompts and individual instructions separately.
| Check | Result |
|---|---|
| BFCL · non live | 12.94% |
| BFCL · live | 9.25% |
| BFCL · irrelevance | 72.66% |
| BFCL · multi turn | 0.00% |
| BFCL · memory | 1.29% |
| BFCL · live relevance | 37.50% |
| BFCL · format sensitivity | 0.25% |
| API-Bank · correct calls | 0/744 |
| IFEval · strict prompt | 133/541 |
| IFEval · strict instruction | 311/834 |
| IFEval · loose prompt | 138/541 |
| IFEval · loose instruction | 319/834 |
These results describe this artifact under the listed protocols. The small tool-use and English sets are research diagnostics.
- Downloads last month
- 40
2-bit