--- library_name: transformers base_model: Qwen/Qwen3.5-4B-Base tags: - code - agent - tool-use - exploration - swe-bench - qwen3.5 language: - en license: apache-2.0 pipeline_tag: text-generation --- > **Building your own AI?** [Join the AI Owner's Club on Skool](https://www.skool.com/open-source-ai-builders-club/about) — a community that owns its models instead of renting them. # Amaara Code Explore A 4B code **exploration** agent. Give it a bug report or a question about a repository and it works the repo with shell and file reads until it locates the relevant code — then stops. It does not modify files. That is deliberate, and measured. Built by [The A-Tech Corporation](https://theatechcorporation.com). Base model `Qwen/Qwen3.5-4B-Base`, two-stage LoRA SFT. ## What it is measured to do Evaluated on **10 SWE-bench Lite (dev) instances at 262,144 context**, 60-step budget, one run each. Metric: did the agent touch a file named in the instance's gold patch? | metric | result | |---|---| | **gold-file recall** (located the file) | **9 / 10 (90%)** | | **median tool calls to first reach it** | **2** | | mean tool calls to first reach it | 2.9 | | **episodes that MODIFIED a gold file** | **7 / 10** | | source-file writes (sed -i, redirects) | 23 | | episodes that submitted a patch | 2 / 10 | **It edits through the shell.** The evaluation scaffold exposes only a `bash` tool, so every modification is a `sed -i`, a redirect, or a heredoc — not a structured edit call. For example, on `pvlib__pvlib-python-1606` it ran `sed -i 's/np\.Inf/np.inf/g' ./pvlib/singlediode.py`, which is one of the gold-patch files. An earlier version of this card reported "0 edit calls" and concluded the model never modifies code. That was an instrument error: it counted calls to an `edit` tool the harness never provides. The model does edit files. Per instance: | instance | tool calls | first hit at | |---|---|---| | marshmallow-code__marshmallow-1343 | 60 | 2 | | marshmallow-code__marshmallow-1359 | 59 | 2 | | pvlib__pvlib-python-1072 | 60 | 2 | | pvlib__pvlib-python-1606 | 49 | 2 | | pvlib__pvlib-python-1707 | 60 | 2 | | sqlfluff__sqlfluff-1517 | 58 | 2 | | sqlfluff__sqlfluff-1625 | 59 | 3 | | sqlfluff__sqlfluff-1733 | 60 | never | | sqlfluff__sqlfluff-1763 | 59 | 8 | | sqlfluff__sqlfluff-2419 | 27 | 3 | ## What it is NOT **Its editing is coarse.** It modifies files with shell redirection and `sed -i` rather than surgical replacements, and 3 of 10 episodes never touched a gold file. It also writes a lot of scratch files (106 in these runs) — reproduction scripts and ad-hoc tests — so a working directory will accumulate junk. Treat it as a capable explorer that will attempt fixes roughly 70% of the time, not as a reliable patch generator. **Seven of ten episodes exhausted the 60-step budget** without submitting. It explores long and terminates late. Budget for that in any harness. **It was trained at 8,192 tokens** but served here at 262,144. The longer window is real (llama.cpp reports `n_ctx_slot = 262144` and an independent 67,243-token needle test passed), but quality beyond the trained length is unverified — the window prevents truncation, it does not guarantee comprehension. **Ten instances is a small sample.** 90% recall on n=10 carries wide error bars. Treat it as a strong indication, not a headline benchmark result. ## Reproduce the measurement The evaluation used [mini-swe-agent](https://github.com/SWE-agent/mini-swe-agent) against a local `llama-server`: ```bash llama-server -m amaara-code-explore-Q8_0.gguf \ -ngl 99 -c 262144 --parallel 1 \ -ub 512 -b 1024 --jinja --host 0.0.0.0 \ --cache-type-k q8_0 --cache-type-v q8_0 ``` `--parallel 1` matters: `-c` is a shared pool, so with the default 4 slots a 262k setting silently gives each request 65k. ## Training | stage | steps | context | notes | |---|---|---|---| | 1 | 3,000 | 4,096 | harness + agent traces | | 2 | 2,000 | 8,192 | continued from stage 1; 28.2% of loss tokens from generalised Paperclip agent traces | Stage 2 was aimed at a specific defect: the stage-1 model repeated identical tool calls in 65% of episodes. After stage 2 the duplicate rate fell to 26%, and after raising the serving context from 32k to 98k one instance fell from 41.8% to 7.0% — part of the original figure was truncation, not looping. Training data was generalised before use: client names, paths, personal names and task codes were substituted with stable pseudonyms, and 25 banned patterns were verified absent from the final corpus. ## Quants | file | size | notes | |---|---|---| | `amaara-code-explore-F16.gguf` | 8.1 GB | reference | | `amaara-code-explore-Q8_0.gguf` | 4.3 GB | the configuration measured above | | `amaara-code-explore-Q6_K.gguf` | 3.4 GB | | | `amaara-code-explore-Q5_K_M.gguf` | 3.0 GB | | | `amaara-code-explore-Q4_K_M.gguf` | 2.6 GB | smallest | Quantisation damage has **not** been measured for this model. The numbers in this card were taken on Q8_0; if you use Q4_K_M, re-measure before relying on them. ## Licence Apache-2.0, inherited from the Qwen3.5-4B base. --- ## Join the AI Owner's Club We build these models in the open with a community of people who want to **own** their AI instead of renting it. If you're training your own models, running them on your own hardware, and refusing to accept that you have to pay per token forever, you'll fit right in. AI Owner's Club **[Join the AI Owner's Club on Skool](https://www.skool.com/open-source-ai-builders-club/about)**